Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
HAL Id: tel-00744666https://tel.archives-ouvertes.fr/tel-00744666
Submitted on 23 Oct 2012
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
Lacunes chargées, étude dans des nano-agrégats desiliciumArpan Deb
To cite this version:Arpan Deb. Lacunes chargées, étude dans des nano-agrégats de silicium. Autre [cond-mat.other].Université de Grenoble, 2012. Français. NNT : 2012GRENY027. tel-00744666
ABCDEE
ABCDAEFBFBEC
FADBAABCDEFFBFCEAAABA
D
EAACDCFA
EAACDCF
DDAEABCA
EDFFFABCDEFFBFCE
EFB FFB!FEFBFB" F BFBCCCE
BA!F#B$%&$"EDB#$E
'CB(F)%D
*EA)*%D
+F,-AE
*EC).&
CF&
CFA&'
To my mother Mrs. Chhabi Deb.
I do not put my faith in any new institutions, but in the individuals all overthe world who think clearly, feel nobly and act rightly. They are the
channels of moral truth. ... Rabindranath Tagore
1
Contents
1 Introduction 61.1 General Ideas . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.2 Types of defects . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.1 Point defects . . . . . . . . . . . . . . . . . . . . . . . 81.2.2 Linear defects . . . . . . . . . . . . . . . . . . . . . . . 101.2.3 Planar defects . . . . . . . . . . . . . . . . . . . . . . . 11
1.3 A brief history of defects in bulk Si . . . . . . . . . . . . . . . 121.4 The Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2 Some Theoretical Concepts : DFT and its components 172.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.2 Density functional theory (DFT) (a brief overview) . . . . . . 17
2.2.1 Hohenberg-Kohn theorems . . . . . . . . . . . . . . . . 182.2.2 The Kohn-Sham (KS) equations . . . . . . . . . . . . . 20
2.3 Different Basis Set Approaches . . . . . . . . . . . . . . . . . . 242.3.1 Atomic Basis Sets . . . . . . . . . . . . . . . . . . . . . 242.3.2 Plane Wave Basis Sets . . . . . . . . . . . . . . . . . . 262.3.3 Wavelet Basis Sets . . . . . . . . . . . . . . . . . . . . 28
2.4 Approximations to the Exchange-Correlation Potential . . . . 302.4.1 The Local Density Approximation (LDA) . . . . . . . . 302.4.2 Overview of the performance of LDA . . . . . . . . . . 312.4.3 Generalized Gradient Approximations (GGA) . . . . . 332.4.4 Meta-GGA . . . . . . . . . . . . . . . . . . . . . . . . 352.4.5 Hybrid Schemes : Combination of Hartree-Fock and
DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352.5 Pseudopotentials . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5.1 Normconserving Pseudopotentials . . . . . . . . . . . . 392.5.2 Fully Nonlocal Pseudopotentials . . . . . . . . . . . . . 412.5.3 Vanderbilt Ultrasoft Pseudopotentials . . . . . . . . . . 42
2
3 State of the Art: Experiments and Theory 453.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453.2 Experimental Methods To Study Point Defects . . . . . . . . . 463.3 Theoretical Methods To Study Point Defects . . . . . . . . . . 50
3.3.1 Treatment of Makov and Payne . . . . . . . . . . . . . 503.3.2 Treatment of Schultz . . . . . . . . . . . . . . . . . . . 553.3.3 Treatment of Freysoldt, Neugebauer and Van de Walle 593.3.4 Treatment of Dabo et.al. . . . . . . . . . . . . . . . . . 61
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4 Hartwigsen-Goedeker-Hutter(HGH) pseudopotentials with NonLinear Core Correction (NLCC) 694.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694.2 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2.1 Non Linear Core Correction . . . . . . . . . . . . . . . 714.3 Tests for the pseudopotentials . . . . . . . . . . . . . . . . . . 77
4.3.1 Hund’s rule . . . . . . . . . . . . . . . . . . . . . . . . 784.3.2 Calculational parameters . . . . . . . . . . . . . . . . . 794.3.3 Atomization energy of molecules using PBE . . . . . . 794.3.4 Atomization energy of molecules using PBE with NLCC 864.3.5 Similar effort with Linear Spin Density Approxima-
tion(LSDA) . . . . . . . . . . . . . . . . . . . . . . . . 944.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5 Charged Defects in Silicon Nanoclusters 985.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 985.2 Comparison of formulae (PBC vs. FBC) . . . . . . . . . . . . 102
5.2.1 PBC formalism to calculate the formation energy . . . 1025.2.2 FBC formalism to calculate the formation energy . . . 104
5.3 Density analysis of the defect systems . . . . . . . . . . . . . . 1055.3.1 Uncharged clusters . . . . . . . . . . . . . . . . . . . . 1055.3.2 Charged clusters . . . . . . . . . . . . . . . . . . . . . 108
5.4 Comparison of results obtained . . . . . . . . . . . . . . . . . 1175.5 Electrostatic model . . . . . . . . . . . . . . . . . . . . . . . . 118
5.5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . 1195.5.2 Details . . . . . . . . . . . . . . . . . . . . . . . . . . . 1205.5.3 Modification . . . . . . . . . . . . . . . . . . . . . . . . 1215.5.4 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.6 Geometry of defects in nanoclusters . . . . . . . . . . . . . . . 1225.6.1 Jahn-Teller Distorsion (JT) . . . . . . . . . . . . . . . 1225.6.2 Jahn-Teller Distorsion in the uncharged nanoclusters . 124
3
5.6.3 Jahn-Teller Distorsion in the charged nanoclusters . . . 1275.7 A simple force model to simulate the JT distorsion . . . . . . 1295.8 The combination of the electrostatic and simple force model . 1305.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1315.10 Future perspectives . . . . . . . . . . . . . . . . . . . . . . . . 132
5.10.1 Different other defects . . . . . . . . . . . . . . . . . . 1325.10.2 Migration energies . . . . . . . . . . . . . . . . . . . . 1335.10.3 Homothetic clusters . . . . . . . . . . . . . . . . . . . . 134
6 Final words and Perspectives 136
4
Chapitre 1
Ce chapitre traite des elements d’introduction a cette these. J’y in-
troduits la description des divers types de defauts presents dans un solide.
Certains resultats et leur importance sont aussi mentionnes. Une rapide com-
paraison avec les methodes existantes pour etudier les defauts est presente.
Les defauts decrits sont les defauts ponctuels, les defauts plans, les disloca-
tions, avec ou sans charge. Une rapide histoire de l’etude des defauts est
aussi mentionnee pour completer l’etat de l’art du domaine. Enfin, certaines
experiences portant sur les defauts sont rappelees.
5
Chapter 1
Introduction
1.1 General Ideas
Silicon (Si), the most common metalloid and the eight most common ele-
ment in the universe by mass is one among the most important material for
industrial use. It is the principal component of most semiconductor devices,
most importantly integrated circuits or microchips. Silicon is widely used
as a semiconductor because it remains more of a semiconductor at higher
temperatures than another material of the same family, germanium, and
also because its native oxide is easily grown in a furnace and forms a good
semiconductor/dielectric interface. In the form of silica and silicates, sili-
con comprises useful glasses, cements, and ceramics. It is also a constituent
of silicones, a class-name for various synthetic plastic substances made of
silicon, oxygen, carbon and hydrogen. In naturally found Si and also for
the industrial uses one often comes across defects in crystalline structure of
the material and hence the study to know more about the pros and cons of
those defects came into practice. There are several phenomena such as the
electrical transport and phototransport, the light absorption and emission
etc., for which the investigation of defects is essential. Sometimes the de-
fects are useful and sometimes they have to be avoided because they make
troubles. In bulk semiconductors, there are various specific applications like
the radiation detectors where the defects are very useful. Since a long time
6
Silicon is used in radiation detectors and electronic devices. Nowadays, these
devices work on submicron technology and they are parts of integrated cir-
cuits with large to very large scale integration. Silicon and silicon based
devices are heavily worked upon in great theoretical and experimental de-
tail, in many fields of physics including particle physics experiments, nuclear
medicine, reactors and space. Defects in the material represent a limiting
factor in the operation of devices. In spite of the huge research being put
into by the scientific community till date, there are a lot of aspects not clari-
fied, related to the behavior of impurities, defects, vacancies in Silicon. And
hence a global understanding of their local structure and properties became
increasingly important due to the reduction in chip sizes and to the increase
of the operation speed. The study of effects of point defects on electronic,
structural and vibration properties of bulk semiconductors, and also on low
size semiconductor structures is a thematic of high interest. This study will
bring contributions to the fabrication of Si based smart materials for photon-
ics, MOS and nanowires based devices etc. The fabrication of new materials
will open up new horizons and enable path-breaking advances in science and
technology.
In bulk semiconductors various defects (sometimes called trap defects), like
the point defects, impurities or local stresses, are located in the volume of
the material. In nanocrystalline semiconductors the trapping phenomena are
dominated by the traps located at the surface or interface, due to the very big
area/volume ratio (of the order of 108m–1 for nanocrystals). These traps are
produced by the adsorption, dangling bonds, and the internal stresses (in-
duced by misfit)[1]. Experimental methods like the photo-induced current
transient spectroscopy (PICTS), the thermally stimulated currents (TSC),
the thermally stimulated depolarization currents (TSDC), and the optical
charging spectroscopy (OCS) are in practice to study those defects. In the-
ory as well scientists are making headway through the calculations of the
stability, energetics of formation, migration, electrostatics etc. of all such
defects. Before going into the further discussion about the defects and their
respective analyses, one may feel the need to classify broadly the types of
defects that are common in Silicon.
7
1.2 Types of defects
In reality there is no existence of a perfect crystal. All crystals have some
defects. These defects do in general contribute to various mechanical and
electronic properties of the material. In fact a part of modern electronic
engineering is based on the manipulation of these defects for efficient use of
the material. Grossly these defect types in silicon crystals can be grouped as
following :
• Point defects. (vacancies, interstitials, impurities etc.)
• Linear defects (dislocations etc.).
• Planar defects (grain boundaries, stacking faults etc.)
1.2.1 Point defects
When in the lattice structure of a material, an atom is missing or is in an
irregular place, then that material is said to have point defects. This class
of defects includes self interstitial atoms, interstitial impurity atoms, substi-
tutional impurities and vacancies. A self interstitial atom is an extra atom
that has crowded its way into an interstitial void in the crystal structure.
A substitutional impurity atom is an atom of a different type than the bulk
atoms, which has replaced one of the bulk atoms in the lattice. Interstitial
impurity atoms are much smaller than the atoms in the bulk matrix and
they are found to fit in the inter-atomic space between the bulk atoms of the
lattice structure. Vacancies are basically the empty spaces where an atom
should be, but is missing. They are common in metals and semiconductors,
especially at high temperatures when atoms are found to change their posi-
tions pretty often leaving behind empty lattice sites. In most cases diffusion
(mass transport by atomic motion) can only occur because of vacancies. The
point defects are also responsible for the lattice strain in the crystal because
of the deformation in the geometry around the defect itself. In the adjoining
Figure 1.1, an example of the occurence of all the point defects discussed
here, is schematically described. The vacancies can be charged giving rise
8
Figure 1.1: Various point defects in the bulk material.
to a special sub-branch of defects known as the charged defects. Even doping
the host material with foreign atoms (technically impurities) produce charged
defects. The cases of doped Silicon crystals can be considered as an example,
where donor atoms provide excess electrons to form n-type silicon and accep-
tor atoms provide a deficiency of electrons to form p-type silicon. It is also
worth mentioning here that combination of the point defects, eg. a vacany
+ an interstitial can give rise to another category of charged defects. Those
defects can influence the formation of cluster defects in the bulk material.
In this regard it is worth mentioning a special type known as the Frenkel
defects. A Frenkel defect, Frenkel pair, or Frenkel disorder is a type of point
defect in a crystal lattice where the defect forms when an atom or cation
leaves its place in the lattice, creating a vacancy, and becomes an interstitial
by lodging in a nearby location not usually occupied by an atom. Frenkel
defects occur due to thermal vibrations, in principle there will be no such
defects in a crystal at 0 K.
9
1.2.2 Linear defects
These defects are found in the material when an array of atoms is displaced
in the crystal structure. The presence of such dislocations affect the physcial
properties of the material to a pretty large extent, in fact the motion of the
dislocations are the cause for the plastic deformations to occur in the ma-
terial. The dislocations can be broadly devided into Edge dislocations and
Screw dislocations. An edge dislocation is said to have occured where an
extra half-plane of atoms is introduced half way in the middle through the
crystal, disrupting the symmetry of nearby planes of atoms. The disloca-
tion is called a line defect because the locus of defective points produced
in the lattice by the dislocation lie along a line. The inter-atomic bonds
are significantly distorted only in the immediate vicinity of the dislocation
line. The screw dislocation is not very straightforward to understand. The
motion of a screw dislocation is also a result of shear stress, but the de-
fect line movement is perpendicular to direction of the stress and the atom
displacement, rather than parallel. Figure 1.2 describes the two different
dislocations schematically. The edge dislocations allow deformation to occur
Figure 1.2: Edge and Screw dislocations.
at much lower stress than in a perfect crystal. One can understand that
10
fact by studying the movement of the edge dislocation. Dislocation motion
is analogous to movement of a caterpillar. A schematic diagram in Figure
1.3 shows the movement of the dislocation. As for the screw dislocations the
Figure 1.3: The movement of the edge dislocation.
dislocations move along the densest planes of atoms in a material, because
the stress needed to move the dislocation increases with the spacing between
the planes. FCC and BCC metals have many dense planes, so dislocations
move relatively easy and these materials have high ductility.
1.2.3 Planar defects
Defects like the Stacking faults, Twin boundaries, Grain boundaries are ex-
amples of planar defects. A stacking fault is a one or two layer interruption
in the stacking sequence of atom planes. Stacking faults occur in a number
of crystal structures. For example we can consider the faults of hcp and fcc
structures. Here the first two layers arrange themselves identically, and are
said to have an AB arrangement. If the third layer is placed so that its atoms
are directly above those of the first (A) layer, the stacking will be ABA. This
indeed is the hcp structure, and it continues ABABABAB. However it is
possible for the third layer atoms to arrange themselves so that they are in
line with the first layer to produce an ABC arrangement which is that of
the fcc structure. So, if the hcp structure is going along as ABABAB and
suddenly switches to ABABABCABAB, there is a stacking fault present.
Another type of planar defect is the grain boundary. The interface between
11
Figure 1.4: (a) Photograph showing a model of the ideal packing in the ABFCC structure, (b) Photograph showing a model of the AB FCC structurewith a stacking fault (c) Differently oriented crystallites in a polycrystallinematerial forming boundaries.
two grains, or crystallites, in a polycrystalline material is known as the grain
boundary. Grain boundaries are such defects in the crystal structure which
are found to decrease the electrical and thermal conductivity of the material.
Grain boundaries limit the lengths and motions of dislocations. It is known
now that having smaller grains (more grain boundary surface area) strength-
ens a material. Different experimental procedures can be used to control the
size of the grains.
1.3 A brief history of defects in bulk Si
Let us now focus on the study of defects in crystalline Silicon. Various exper-
iments and theoretical calculations have revealed some interesting properties
of the defects. In both forms of study there are frontiers which are difficult to
conquer, mostly because of the complexity of the systems. In some cases the
experiments are very difficult to carry out, whereas in some cases to construct
the real theoretical picture of the system is extremely challenging. Some facts
which will be mentioned in the following section are well researched and es-
tablished. As for example we have known that the stability of crystalline
silicon comes from the fact that each silicon atom can accommodate its four
valence electrons in four covalent bonds with its four neighbors. The pro-
12
duction of primary defects or the existence of impurities or defects destroys
this fourfold coordination. It has been established[3][4] that the structural
characteristics of the classical vacancy in bulk Si are: the bond length in
the bulk is 2.35 A and the bond angle –109 degrees . The formation energy
is 3.01 eV (p-type silicon), 3.17 eV (intrinsic), 3.14 eV (n-type). For inter-
stitials, different theoretical works have reported the possibility of various
structural configurations. They are a) the hexagonal configuration, a sixfold
coordinated defect with bonds of length 2.36 A, joining it to six neighbors
which are fivefold coordinated; b) the tetrahedral interstitial is fourfold co-
ordinated; has bonds of length 2.44 A, joining it to its four neighbors, which
are therefore fivefold coordinated; c) the split − < 110 > configuration: two
atoms forming the defect are fourfold coordinated, and two of the surround-
ing atoms are fivefold coordinated; d) the ‘caged’ interstitial contains two
normal bonds, of length of 2.32 A, five longer bonds in the range 2.55 2.82
A, and three unbounded neighbors at 3.10 3.35 A. The calculations [6]-[8]
found that the tetrahedral interstitial and caged interstitial are metastable.
For interstitials, the lowest formation energies (calculated theoretically) in
eV are 2.80 (for p-type material), 2.98 (for n-type) and 3.31 in the intrinsic
case respectively. Authors have reported that in silicon the vacancy takes on
five different charge states in the band gap, viz. V 2+, V +, V 0, V − and V 2–[9].
But about the charge states of the self-interstitial there are differences of
opinions in reports by Lopez[10] et. al. The experimental examination of
primary point defects buried in the bulk is difficult and for various defects
this is usually indirect. In a series of theoretical studies[2] and correlated
EPR and DLTS experiments of Watkins and co-workers[4], it became possi-
ble to solve some problems associated with the electrical level structure of the
vacancy. In crystalline silicon bombarded with energetic projectiles, the di-
vacancy center is being studied for quite some time now by numerous authors
applying various experimental techniques, e.g. EPR [3], photoconductivity
[11], infrared absorption [12], electron-nuclear double resonance (ENDOR)
or deep level transient spectroscopy (DLTS) and at the room temperature it
is considered as a stable defect. The undisturbed configuration of divacancy
could be viewed as two vacant nearest neighbor lattice sites. Experiments by
13
Watkins et. al. with isochronal and isothermal annealing studies reveals that
the divacancy anneals out at 570 K[3]. Since then many other experiments
have tried to comment on the formation mechanism and charge states of the
divacancy.
In 2002, in an important report, S. Goedecker, T. Deutsch and L. Billard
[13] predicted the existence of a new type of primary defect in silicon and
thus a new type of symmetry in the crystallography of the material. Their
work proved the existence of a fourfold coordinated defect which is found to
be more stable. Experiments by Lazanu and Lazanu [14] proved the exis-
tence of such defects. Lattice Monte Carlo calculations are done by Damien
Caliste and Pascal Pochet[15] to study the diffusion of vacancies in bulk Si.
They have reported the results of simulations of vacancy-assisted diffusion
in silicon to show that the observed temperature dependence for vacancy
migration energy is explained by the existence of three diffusion regimes for
divacancies. The geometry and energetics of the normal divacancies and split
divacancies are discussed as well. In a follow up work Damien et. al. have
presented an analysis of stress-enhanced vacancy-mediated diffusion in biax-
ially deformed Si (100) films as measured by the strain derivative (Q′) of the
activation energy[16].
1.4 The Motivation
There are in fact two parts of this thesis. The first part deals with the
accuracy of Hartwigsen-Goedeker-Hutter (HGH) pseudopotentials with Non
Linear Core Correction (NLCC). The main idea is to use the pseudopoten-
tials for further accurate calculations. One can argue of course the validity
of such a job. We have tried to find out a less complex approach with
the norm-conserving pseudopotentials to produce as accurate results as that
of the Projector Augmented Wave[80] (PAW) approach or that of the all-
electron calculations. The success of our approach would ensure a method
other than the PAW method for accurate calculations of various physical
quantities. This is an attempt to correct the otherwise implied linearization
of the pseudopotential. Chapter 4 would deal in detail with this part of the
14
research. For the next and more important part of the research we wish
to concentrate ourselves in the charged point defects in silicon nanoclusters.
Up to now, as will be explained in Chapter 3, the charged defects have been
simulated within the framework of Periodic Boundary Conditions (PBC).
Indeed this PBC approach is good to simulate an infinite bulk. But, as will
be shown in detail in the following chapters, this PBC approach comes with
its own artifacts which can be difficult to deal with. The electrostatics, in
particular for those systems is not simple to represent and a lot of mathe-
matical jargon is needed to account for some other unwanted interactions.
In our approach we have tried to find a method with which we can deal with
isolated charged systems and also can comment on bulk properties even when
we are working with Free Boundary Conditions (FBC). Our motivation is to
find the correct electrostatics of the charged defects in the nanoclusters and
its ability to extrapolate those results for the infinite bulk cases. Chapter 5
deals with this part of the research in detail. Before going into the details
of the current research work it is handy to have an overview of some basic
theoretical concepts.
15
Chapitre 2
Ce chapitre traite des differents concepts theoriques a la base de cette
these. Une vue generale de la theorie de la fonctionnelle de la densite (DFT)
dans ces differents aspects est decrite. J’aborde aussi les differentes fonction-
nelles d’echange-correlation (ECF) d’interet pour la these, ainsi que la theorie
sous-jacente. La these utilisera par la suite une fonctionnelle GGA. Enfin,
je traite aussi les differentes bases existantes pour resoudre les equations
de Kohn-Sham (KS), montrant ainsi l’interet des ondelettes pour le traite-
ment des defauts. Tous les calculs ayant ete fait avec l’approximation des
pseudo-potentiels, la theorie est abordee ici. La these est basee sur la forme
des pseudo-potentiels de Hartwigsen-Goedecker-Hutter (HGH) pour le calcul
des defauts aborde au chapitre 5, et les HGH avec correction non-lineaire de
cœur pour le calcul des energies d’atomisation decrites dans le chapitre 4.
16
Chapter 2
Some Theoretical Concepts :
DFT and its components
2.1 Introduction
In this chapter we discuss in order of appearance the basic theoretical con-
cepts of Density Functional Theory, along with its major components - the
Pseudopotentials and the Exchange Correlation Functional. The different
basis sets which can be used to solve the Kohn-Sham equations are also
described with their respective pros and cons.
2.2 Density functional theory (DFT) (a brief
overview)
Since its discovery almost four decades back, DFT, a theory of electronic
ground state structure, has been one among the most important tools to
understand the ground state properties of molecules, solids and clusters.
Theoretical physicists and chemists find in it an altrenative to the traditional
methods of quantum chemistry dealing with many-electron wave-function.
DFT in its full glory is a thematic continuation of the previous approximate
methods although itself being exact in principle. With a motive to reduce the
number of parameters needed to describe the many body system, DFT works
17
centering around the electron density ρ(r) of the electronic system. The main
goal of almost all physical systems is to know its various physical properties
and structure, and hence calculation of the total energy of the system is most
important. Keeping in accordance with that goal, all the contributions to the
total energy are expressed in terms of the electron density in the framework
of DFT. In the following subsections various important structural parts of
the DFT formalism are explained.
2.2.1 Hohenberg-Kohn theorems
Among the structural pillars of DFT Hohenberg-Kohn (HK) theorems [17]
are of foremost importance. The first HK theorem proves that the ground
state electron density ρ is sufficient to determine , in principle, the total
energy of the system and hence any ground state property of the system
without the knowledge of the many electron wave function can thus be ob-
tained. The second theorem states that the ground state energy of the system
is the minimal value of the energy functional of the electronic system. Both
the two theorems are proven here in the following subsections.
First Hohenberg-Kohn theorem. The ground state density ρ (r) of a
system of many interacting electrons moving in some external potential v(r)
determines this potential uniquely.
Proof. [18] By reductivo ad absurdum. Let us consider two external potentials
v1(r) and v2(r) having the same ground state electron density ρ(r), with
Hamiltonians H1 and H2 respectively, and non-degenerate ground states |Ψ1〉and |Ψ2〉, therefore
H1 |Ψ1〉 = E01 |Ψ1〉 (2.1)
H2 |Ψ2〉 = E02 |Ψ2〉 . (2.2)
Let us apply the variational principle at this point
E01 < 〈Ψ2|H1 |Ψ2〉 = 〈Ψ2|H2 |Ψ2〉+ 〈Ψ2|H1 −H2 |Ψ2〉 (2.3)
< E02 +
∫ρ(r) [v1(r)− v2(r)] dr . (2.4)
18
Evidently by interchanging subscripts 1 and 2 in the above inequality another
inequality can be formed of similar structure. By adding both inequalities
we can get
E01 + E0
2 < E02 + E0
1 (2.5)
which is obviously not true. Hence by contradiction to our initial assumption
we can conclude that there is one and only one external potential
associated with the ground state electronic density.
Since ρ(r) is known for the system, all other ground state properties like
the particle number, and the external potential, the total energy E [ρ], the
kinetic energy T [ρ], the potential energy V [ρ], etc. can be calculated directly.
Second Hohenberg-Kohn theorem. If any electron density, given by,
ρ(r) is such that∫ρ(r)dr = N and ρ(r) > 0 then
E0 6 Ev [ρ] = T [ρ] + Vne [ρ] + Vee [ρ] ≡∫
v(r)ρ(r) dr+ FHK [ρ] . (2.6)
where E0 is the ground state energy, Vne and Vee are potentials due
to nucleus-electron and electron-electron interactions. FHK is the univer-
sal functional as it describes a treatment of the kinetic and internal potential
energies which is same for all systems.
Proof. According to the first HK theorem any electron density ρ(r) uniquely
determines the external potential v(r) consequently with its own Hamiltonian
and the ground state wave function 〈Ψ| associated with this Hamiltonian. Let
〈Ψ| describe the ground state of a system with density ρ and Hamiltonian
H. According to the Rayleigh-Ritz variational principle, for any electronic
state 〈Ψ| ⟨Ψ∣∣∣H
∣∣∣Ψ⟩≥ 〈Ψ|H |Ψ〉 = E [ρ] (2.7)
and ⟨Ψ∣∣∣H
∣∣∣Ψ⟩=
∫v(r)ρ(r) d3r+ T [ρ] + Vee [ρ] = Ev [ρ] . (2.8)
Therefore
Ev [ρ] ≥ Ev [ρ] (2.9)
19
2.2.2 The Kohn-Sham (KS) equations
In the framework of DFT KS equations [18] are essentially Schroedinger
equations of a system of noninteracting particles that would generate the
same density as one given system of interacting particles. To that end we
need to derive the equations for the single particle system that yields the
density ρ(r). To start with let us focus in deriving the equations for a system
of noninteracting particles. For a noninteracting system, the total energy of
the system can be written as
E[ρ] = Ts [ρ] +
∫v(r)ρ(r) , (2.10)
where v is the net potential and Ts [ρ] is the kinetic energy of the ground
state of the system of noninteracting electrons having a density ρ(r). The
density of the system is given by
ρ(r) =N∑
i=1
|ϕi(r)|2 . (2.11)
The number of occupied states N being constant, let us consider the variation
in density as
δρ(r) =N∑
i=1
δϕi∗(r)ϕi(r) (2.12)
Since the density in equation 2.11 is stationary with respect to the given
variation (the total number of particles being constant), the integral of the
variation of density is evidently zero.
∫δρ(r)dr =
∫ N∑
i=1
δϕi∗(r)ϕi(r)dr = 0 . (2.13)
Let us now apply the trick of Lagrange multiplier. For this case with the
lagrangian multiplier ǫ, the variation in the energy functional takes the form
20
as
δE[ρ(r)] =
∫δρ(r)
v(r) +
δ
δρ(r)Ts[ρ(r)]− ǫ
dr = 0 . (2.14)
The system we have at hand is of noninteracting particles. Hence we can
write the ground state energy eigenvalue of the system as
E =N∑
i=1
ǫi (2.15)
and obtain as single-particle equations
[−1
2∇2 + v(r)− ǫi
]ϕi(r) = 0 (2.16)
The objective is to find equations for the states ϕi, but it must be kept in
mind that these equations (2.16) are valid only for noninteracting particles.
Obviously they yield the density ˜ρ(r) of a system of noninteracting particles.
Whereas in reality the quantity we are interested in is the density ρ(r) of
a system of interacting particles. In the framework of DFT equation (2.10)
needs to be rewritten as
E[ρ] = FHK [ρ] +
∫vext(r)ρ(r) , (2.17)
where vext is the external potential and FHK [ρ] is the density functional,
which we can split into expressions like
FHK [ρ] = Ts [ρ] +1
2
∫ ∫ρ(r)ρ(r′)
|r− r′| drdr′ + EXC [ρ] , (2.18)
In 2.18, T [ρ] is the noninteracting part of the kinetic energy of this many
electron system we are dealing with
Ts [ρ] =∑
i
〈ϕi| −1
2∇2
r|ϕi〉 . (2.19)
The second term in equation 2.18 describes the classical electrostatic inter-
action between the noninteracting particles. This is the Hartree energy of
21
the system. EXC [ρ] is the exchange-correlation energy, which takes in ac-
count the entire electron-electron interaction effect beyond the Hartree term
viz. the exchange energy due to the Pauli exclusion principle, the correla-
tion energy and the difference in kinetic energy between an interacting and a
noninteracting system. In practice the second and third term are collectively
considered with the external potential as the effective Kohn-Sham potential
veff (r, ρ(r)) = vext(r) +
∫ρ(r′)
|r− r′|dr′ +
δEXC [ρ(r)]
δρ(r). (2.20)
The exchange-correlation potential is described by the third term as
vxc(r) =δExc [ρ]
δρ. (2.21)
Hence equation 2.17 can alternatively be written as
E[ρ] = Ts [ρ] +
∫veff (r)ρ(r) , (2.22)
The particles described by equation 2.22 and 2.10 are of course different yet
since those two equations are essentially similar they will yield similar single
particle equations. Finally we can write the most appropriate form of the
Kohn-Sham equations as
[−1
2∇2 + veff (r)
]ϕi = ǫiϕi , (2.23)
ϕi and ǫi being the single particle orbitals and energies. At this point
it is pretty clear that the ground state electron density can be computed
from the self-consistent solution of the KS equations as in 2.23 only if the
expression for the exchange-correlation potential is known. In practice, the
density is calculated with the formula in equation 2.24 and consequently
the obtained density is inserted in the exchange-correlation potential as in
equation 2.21. Then the new eigenstates and new electron density of the
22
system are calculated, until the required convergence is reached.
ρ(r) =N∑
i=1
|ϕi(r)|2 (2.24)
It is noteworthy here that the KS equations as in equation 2.23 yield eigen-
states ϕi of fictitious particles which only have the same density as the real
particles. Hence it must be kept in mind that the ground state energy of the
real system is not the simple sum of the Kohn-Sham single-particle eigen-
values as for noninteracting particles 2.15. Let us now have a look at the
ground state energy (of the real system). It would look like
E =N∑
i=1
ǫi + Exc[ρ(r)]−∫
vxc(r)ρ(r)dr−1
2
∫ρ(r)ρ(r′)
|r− r′| drdr′. (2.25)
KS DFT is a technique to describe the ground state energy and density, more
to say almost all ground state properties of a many body system in terms
of single particle equations and states. It is trap that one may find himself
in if one believes, that in this formalism, real electrons are being described
as independent particles experiencing an external field of the ions and of
all other electrons. In reality the KS equations are merely the equations of
noninteracting fictitious particles, whose density is the same as the density of
real electrons in the ground state. Hence it can be said that the eigenstates
of the KS equations do not have a direct physical meaning. However, in a
system of N occupied states the N-th KS eigenvalue, ǫN , in equation (2.23)
finds one a physical interpretation [20]
• The ionization potential (IP) of a finite system is given by
ǫN = −IP (2.26)
• And the same is the chemical potential for extended systems µ
ǫN = µ (2.27)
23
It depends on the approximation used for the exchange-correlation func-
tional/potential how close ǫN comes to the exact ionization potential or
chemical potential [21].
2.3 Different Basis Set Approaches
As one can expect there are different approaches to solve the KS equations.
The approaches are based on the basis sets on which this KS equations are
expanded and then solved accordingly. Within this chapter I would discuss
the three main approaches in terms of basis sets viz. the localized basis set
approach, the plane wave approach and the wavelet basis set approach.
2.3.1 Atomic Basis Sets
Since the first efforts to solve the KS equations the use of atomic basis sets
has been serving the cause to a very good effect. In this approach the molec-
ular orbitals (MO) are expressed as a linear combination of atomic orbitals
(LCAO). The basic strength of the LCAO approach is its general applicabil-
ity as in reality it can work on any molecule with any number of atoms. To
take the LCAO concept another step ahead to facilitate the calculations we
can use a larger number of atomic orbitals (AO)(e.g. a hydrogen atom can
have more than one s AO, and some p and d AOs, etc.). This helps us to
achieve a more flexible representation of the MOs and therefore more accu-
rate calculated properties according to the variation principle. One can also
use AOs of a particular mathematical form that simplifies the computations
(but are not necessarily equal to the exact AOs of the isolated atoms). These
AOs are called the basis functions, or more precisely the localized atomic ba-
sis functions. Instead of having to calculate the mathematical form of the
MOs (impossible on a computer) the problem is reduced to determining the
MO expansion coefficients in terms of the basis functions. Functions that
resemble hydrogen AOs (Slater functions in other words) are very suitable
for expanding MOs because they have the correct shape (a) near the nucleus
(shape of a cusp) and also (b)far from the nucleus (decay like exp−ar). The
24
main reason for the preference of the Gaussian functions is the fact that they
allow for efficient computation of molecular integrals. In quantum chem-
Figure 2.1: Illustration of the simplicity of the Gaussian functions over theslater functions
istry terminology a single Gaussian function is called a primitive Gaussian
function, or primitive GTO (Gaussian Type Orbital). Some programs use
cartesian primitive GTOs while the others use spherical primitive GTOs. In
mathematical terms spherical and cartesian functions are the same for up
to l = 1 (p functions) but differ slightly for l = 2 or higher. In practice,
fixed linear combinations of primitive Gaussian functions are used which are
called Contracted Gaussians (CGs). The simplest kind of CGs are the STO-
nG basis sets . These basis sets attempt to approximate Slater-type orbitals
(STOs) by n primitive Gaussians. The STO-nG basis sets are not that sat-
isfactory as they include only one CG per atomic orbital. Improved basis
sets are obtained by including more than one CG per atomic orbital, e.g.:
DZ (“double zeta”), TZ (“triple zeta”), QZ (“quadruple zeta”). Or the im-
provement can also be achieved by the use of one CG per core atomic orbital
and more than one for the valence atomic orbitals, e.g. SV, 3-21G, 4-31G,
6-31G, 6-311G. Increasing the number of CGs per atomic orbital will not
25
actually give a good quality basis set because in reality there are other types
of CGs which are to included. For example one can include CGs of angular
momentum higher than in the valence orbitals of each atom. These polar-
ization functions enhance the flexibility of atoms to form chemical bonds in
any direction and hence improve the calculated molecular structures. Few
examples of such type are DVP, TZP, cc-pVDZ, cc-pVTZ etc. One can also
take in account the CGs which extend further from the nucleus than the
atomic orbitals. Such diffuse functions improve the predicted properties of
species with extended electronic densities such as anions or molecules form-
ing hydrogen bonds. Basis sets are considered balanced when they include
both polarization and diffuse functions. Examples of these types comprises
6-31+G*, 6-311++G**, aug-cc-pVDZ etc. But completeness for these types
of basis sets also calls for it to be infinite which is computationally unachiev-
able. Indeed the atomic basis sets give out the real physics of the molecules
but the two point integrations are exceedingly difficult to calculate with such
basis sets. Again there is a problem of physical convergence of the computed
results with these basis sets. In principle more informations can be put in to
the computational structure but to need the complete picture what we need
is an infinite basis set of this kind. Which of course is not possible to realize
in reality.
2.3.2 Plane Wave Basis Sets
This approach is one where the Fourier representation of the equations bear a
heavy significance. In reality most computational operations are easy to work
with in the Fourier space which hands over an advantage to the plane wave
basis sets in DFT calculations. According to the Bloch’s theorem of solid
state physics, the electronic wave functions at each k point can be expanded
in terms of a discrete plane wave basis set. In principle, an infinite plane-wave
basis set is required to expand the electronic wave functions. However, the
coefficients for the plane waves with small kinetic energy are more important
than those with large kinetic energy. Thus one can understand that there is a
possibility for the plane-wave basis set to be truncated to include only plane
26
waves that have kinetic energies less than some particular cutoff energy. The
basis set would be infinite if a continuum of plane-wave basis states were
required to expand each electronic wave function, irrespective of the small
magnitude of the cutoff energy. Thus one feels a need to make this apparently
infinite basis set to a finite one. Application of the Bloch theorem allows
the electronic wave functions to be expanded in terms of a discrete set of
plane waves. To that end if one introduces a cutoff energy to the discrete
plane-wave basis set we can have a finite basis set. Hence one can say that to
achieve completeness plane wave basis sets should be infinite as well but with
the cutoff approximation appropriately accurate results can be obtained. An
error will be there for sure in the computed total energy due to the truncation
of the plane-wave basis set at a finite cutoff energy. However, it is possible
to reduce the magnitude of the error by increasing the value of the cutoff
energy. In principle, the cutoff energy should be increased until the calculated
total energy has reached convergence. One of the difficulties associated with
the use of plane-wave basis sets is that the number of basis states changes
discontinuously with cutoff energy. Generally these discontinuities are found
to occur at different cutoffs for different k points in the k-point set. Another
important aspect is that at a fixed-energy cutoff, a change in the size or shape
of the unit cell will cause discontinuation in the plane-wave basis set. This
problem is generally taken care of by using denser k-point sets, so that the
weight attached to any particular plane-wave basis state is reduced. However,
the problem is still present even with quite dense k-point samplings. It can
be handled by applying a correction factor which manages to account for
the difference between the counts approximately number of states in a basis
set with infinitely large number of k points and the number of basis states
actually used in the calculation [56]. In thsi approach though one needs to
compute the coefficients which are in general large in number. This aspect
raises a question about the speed of the process for calculations of very large
systems. In reality this is an order N method. Very large systems can be
tackled if we can use the plane wave DFT codes with parallel computation.
But it is very difficult to parallelize the plane wave computations mainly
because of the complexities raised by Fast Fourier Transformation (FFT).
27
2.3.3 Wavelet Basis Sets
A novel approach to solve the KS equations were shown to be very suc-
cessful by Genovese and his colleagues[43]. They have used the Daubechies
wavelets as a powerful systematic basis set for electronic structure calcula-
tions and have shown that they are good because they are orthogonal and
localized both in real and Fourier space. Daubechies wavelets have been
found to have all the properties that a DFT program would like to have of a
basis set used for the simulation of isolated or inhomogeneous systems. They
form a systematic orthogonal and smooth basis, localized both in real and
Fourier spaces and that allows for adaptivity. Hence a DFT approach based
on such functions will be very good to satisfy the requirements of precision
and localization found in many applications. A wavelet basis comprises a
family of functions generated from a mother function and its translations
on the points of a uniform grid of spacing h. Here the number of basis
functions is increased by decreasing the value of h (which is an analog to
the cutoff energy for the case of plane wave basis sets). Because the basis
set in systematic , the numerical description is reported to be more precise.
The degree of smoothness determines the speed with which one converges
to the exact result as h is decreased. The degree of smoothness increases as
one goes to higher order Daubechies wavelets. In the method used for the
DFT code BigDFT (the one which is used for the calculations in this thesis)
Daubechies wavelets of order 16 are used. A very high rate of convergence
is achieved in comparison to other finite difference, finite element, or real
space methods[57][58][59]. Luigi et. al. have discussed the need of the lo-
calization of the basis sets in real space as essential for molecular systems.
This is becaude for basis sets that are not localized in real space are waste-
ful in the context of molecular systems. For example, with plane waves one
has to fill an orthorhombic cell into which the molecule fits. There can be
regions of the cell that may contain no atoms and hence no charge density.
But the plane wave approach cannot be used to utilize this scenario. Since
Daubechies wavelets have a compact support, one can consistently define a
set of localization parameters, which allows the researchers to put the basis
28
functions only on the points that are very close to the atoms. The compu-
tational volume in this method is thus given only by the union of spheres
centered on all the atoms in the system. Real space localization is also nec-
essary for the implementation of linear scaling algorithms. Hence this family
of basis set is a worthy contender for developing such algorithms. A technical
fact can be looked into at this point. For a given system, the convergence
rate of the minimization process depends on the highest eigenvalue of the
Hamiltonian operator. Since the high frequency spectrum of the Hamilto-
nian is dominated by the kinetic energy operator, high kinetic energy basis
functions are therefore also approximate eigenfunctions of the Hamiltonian.
A function localized in Fourier space is an approximate eigenfunction of the
kinetic energy operator. By using such functions as basis functions for the
KS orbitals, the high energy spectrum of the Hamiltonian can thus easily
be preconditioned. It is of course known that a high degree of adaptivity
is necessary for all-electron calculations since highly localized core electrons
require a much higher spatial resolution than the valence wave function away
from the atomic core. This fact is also discussed in the previous paragraphs.
The important fact is that pretty high adaptivity can be obtained with a
wavelet basis. In BigDFT use of pseudopotentials is undertaken to a very
good effect. This is because such pseudopotentials are the easiest way to in-
corporate the relativistic effects that are important for heavy elements. The
use of pseudopotentials readily reduces the need for adaptivity and one has
therefore only two levels of adaptivity. Now one will have a high resolution re-
gion that contains all the chemical bonds and a low resolution region further
away from the atoms where the wave functions would decay exponentially
to zero. In the low resolution region each grid point carries a single basis
function. In the high resolution region it carries in addition seven wavelets.
In terms of degrees of freedom, the high resolution region is therefore eight
times denser than the low resolution region. In comparison with a plane
wave method this wavelet method is therefore particularly efficient for open
structures with large empty spaces and a relatively small bonding region.
29
2.4 Approximations to the Exchange-Correlation
Potential
In this section we talk about the approximation which is inherent to the DFT
methods. It is the approximation to the exchange-correlation potential.
In equation (2.21)the exchange correlation potential is defined as
vxc(r) =δExc [ρ]
δρ. (2.28)
Although the KS equations appear very promising yet without the explicit
knowledge of the exchange correlation functional the KS equations cannot
be solved exactly. Among the cons of DFT finding a good approximation
to the exchange correlation potential is of prime importance. In this section
two very important approaches for the approximation to the potential are
discussed.
2.4.1 The Local Density Approximation (LDA)
In the framework of LDA [19] the electron density of local area of the in-
homogeneous system is approximated with the same density as that of a
homogeneous electron gas in the same extremity. If we take ǫxc to be the
exchange correlation energy per electron of a homogeneous electron gas of
density ρ, then the real exchange correlation energy functional can be ap-
proximated by
Exc[ρ] ≈ ELDAxc [ρ] ≡
∫ρ(r)ǫxc(ρ(r)) dr . (2.29)
the expression for vxc can be obtained by taking the functional derivative of
Exc with respect to ρ
vLDAxc =
δExc
δρ= ǫxc(ρ) + ρ
∂ǫxc∂ρ
(2.30)
30
and, since the system is not homogeneous, ρ = ρ(r). It is assumed hereby
that ǫxc can be calculated accurately for the homogeneous electron gas. In
fact, ǫxc can be divided into exchange and correlation parts ǫxc = ǫx + ǫc.
The exchange part is known analytically and given by
ǫx =3
4
(3ρ(r)
π
)1/3
(2.31)
and the LDA exchange potential is given as
vx(r) = −(3
πρ(r))1/3 , (2.32)
while for the correlation part one has to fall back on the Perdew-Zunger
parametrization of quantum Monte-Carlo data for the electron gas [23]. It is
obvious from the discussion till here that for a homogeneous electron gas, the
LDA exchange-correlation functional is exact. Despite the fact that for most
applications, especially for isolated systems, the electron density differs by a
fair margin from that of a homogeneous electron gas, the approximation still
produces acceptable and good results. This can be understood from the ex-
planation that the LDA satisfies the sum rule which expresses normalization
of the so-called exchange correlation hole [18]. In other words, given that
an electron is at the position r, the electron density for the other electrons
is depleted near r. A ’hole’ in the density distribution ρ(r′). is dug by the
electron at position r due to Pauli principle and electron-electron interaction.
It is thus normalized as
∫ρxc(r, r
′)dr′ = −1 (2.33)
2.4.2 Overview of the performance of LDA
In a report[29] results showing the success of LDA formalism have listed.
In a thumbnail representation, the success story of LDA can be tabulated as
the following,
• Binding energies are often better than 1 eV but in some s–d bonded
31
systems the error can be twice or even three times as large. There is a
systematic over binding.
• Equilibrium distances are generally accurate to within 0.1 A, They are
systematically too short. Hence in general it can be said that geometries
are accurately computed within this framework.[22]
• Vibrational frequencies are accurate to within 10–20 %. There exist
occasional cases with larger errors.
• Charge densities are better than 2 %.
• LDA results are nearly always much better than those of the Hartree–Fock
(HF) approximation.
As for the validity of the above observations we can have a look at the follow-
ing tables in figure 2.2 taken from the works of Muller, Jones and Harris[30].
The results shown here, however, also indicate some problematic deficiencies.
Most notable is the systematic overbinding predicted by the LDA, particu-
larly for the s-d bonded systems. Although the overbinding is to a much
lesser extent yet is reflected in a small but relatively systematic underesti-
mate of the bonding distances. The following small list gives instances of
systems for which the LDA poses some serious problems.
• The transition-metal oxides FeO and CoO are erroneously predicted
to be metallic. But, MnO and NiO come out as anti-ferromagnetic
insulators in accordance with experiment[31].
• Solid Fe is predicted to be an fcc paramagnet [32] but is a bcc ferro-
magnet at low temperatures.
• In many semiconductors, the LDA gives the metal-insulator transition
at much too large volumes[33].
• The LDA predicts the wrong dissociation limits for a large number of
molecules[34].
• The LDA predicts incorrect ground states for many atoms[35].
32
Figure 2.2: Ground-state properties of the molecules H2O, NH3, and CO2,as obtained from the LDA and from experiment. It is assumed that thenumerical errors involved in obtaining the LDA results are negligible in com-parison to the deviation between theory and experiment. The equilibriumdistances in Table II are probably exceptions to the assumptions becausethey do not conform to the general expectations of bond distances being tooshort within the LDA.
• The LDA gives unstable negative atomic ions in many cases when these
are stable[35].
2.4.3 Generalized Gradient Approximations (GGA)
As stated in the previous section, we can understand that although LDA
is quite adequate for some cases, yet for most systems a higher accuracy is
desired. An error in a binding energy of the order of 20 % or 1 eV is, e.g.,
not acceptable in the study of chemical reactions. In this field what would
be great is the binding-energy errors being of the order 0.05 eV or less. The
simplicity of DFT calculational methods as compared to traditional many-
body techniques has, however, spawned considerable efforts to improve on
33
the LDA. Attempts to go beyond the LDA are based either on an improved
description of the exchange-correlation hole in real space or on a description of
exchange-correlation energies in reciprocal space usually leading to so-called
generalized gradient approximations (GGA). In recent years, the largest effort
has gone into the reciprocal-space approach which so far has been the most
successful ab initio DFT method. It is common and intuitive to consider
GGA corrections as some sort of next order corrections to the LDA. In this
framework a functional dependence on the gradient of the density is added
to ǫxc, i.e.,
EGGAxc [ρ] =
∫d3r ρ(r)ǫxc (ρ(r),∇ρ(r)) dr . (2.34)
As reported in [18], compared to the LDA approximation, the error for ion-
Figure 2.3: The errors (in eV) in the binding energies of the first-row dimersas obtained from different density functionals defined below. ∆ is the averageabsolute error for each functional.
ization energies is reduced by factors of 3-5 for the GGA corrections. In the
DFT community one can find several different GGA functionals. An illus-
trative comparison is made in [24]. The table in Figure 2.3 from the article
of Von Barth[29] gives a comparative idea of the accuracy of the corrective
functionals as well. Perdew, Burke, and Ernzerhof[25] have described a func-
tional form (PBE) that has several attractive features. The accuracy of the
34
PBE functional for atoms and molecules has been compared with results
of other popular functionals like LSD, BLYP, and B3LYP[40] by Ernzerhof
and Scuseria[26]. The PBE functional performed as well as B3LYP for the
properties considered by these authors. Generalized gradient approximations
generally lead to improved bond angles, lengths, and energies. In particular,
the strengths of hydrogen bonds and other weak bonds between closed shell
systems are significantly better than local density results. However, the self
interaction problem remains, and some asymptotic requirements for isolated
atoms are not satisfied.
2.4.4 Meta-GGA
Following the development of GGA, the next steps in the gradient approx-
imations were taken with a view to incorporate the kinetic energy density.
A version based on the PBE form was described by Perdew et.al.[28]. This
form includes the kinetic energy density for the occupied Kohn Sham or-
bitals. However, this and other forms initially developed included parame-
ter(s) found by fitting to experimental data. This last feature was avoided in
the recent work of Tao, Perdew, Staroverov and Scuseria[37][36] whose form
satisfied the requirement that the exchange potential be finite at the nucleus
for ground state one and two electron densities. Extensive numerical tests
for atoms, molecules, solids, and jellium surfaces showed generally very good
results.
2.4.5 Hybrid Schemes : Combination of Hartree-Fock
and DFT
In many reports it is mentioned just how poor the exchange energy differ-
ences could be between states whose wave functions have different nodal
structures. It has also been noted for many years[38] that errors in the local
density descriptions of exchange and correlation tend to balance. Probably
the final step in constructing the correct functional is to come up with the
Hybrid Functional. Hybrid functionals are a section of approximations to the
35
functional in DFT that blend a portion of Exact Exchange from Hartree-Fock
(HF) theory with exchange and correlation from other sources (ab initio, such
as LDA, or empirical). This suggests that a combination of Hartree-Fock ex-
change and DFT calculations could be useful as in the following expression,
EhybridXC = αEHF
X + EC , (2.35)
where α can be chosen to satisfy the particular criteria. A formal justifi-
cation for such hybrid schemes was given by Gorling and Levy[39]. Among
the ones which are used a lot these days in the community B3LYP[40] is
one. Hybridization with HF provides a simple scheme for improving many
molecular properties, such as atomization energies, bond lengths and vibra-
tion frequencies. It is to be noted here that those calculations which are done
with the first generation functionals enjoy a further improvement in accuracy.
Enrique R. Batista[41] and his co-workers made a benchmark calculations as
to have a comparative idea about the performance of the exchange correla-
tion functionals. According to their report the Diffusion Monte Carlo (DMC)
methods give the highest accuracy in terms of the calculations of interest.
Hence their comparing standard for all the exchange-correlation functionals
is the set of results from DMC. It goes without saying that DMC is very
expensive as far as computation time is concerned. It is an attempt to sum-
marize the comparative performances of the hybrid pseudopotentials. Ac-
cording to their report[41] the Heyd-Scuseria-Ernzerhof [42](HSE) functional
has been found to be in excellent agreement with their DMC benchmark (Fig-
ure 2.4). A comparative idea has been put into paper between the various
hybrid functionals like Perdew-Wang-91(PW91)[27], Tao-Perdew-Scuseria-
Staroverov [37](TPSS) and the previously mentioned functionals. The com-
parative idea shows that the hybrid functional like HSE significantly improves
the agreement between DFT and DMC (DMC being established as the most
accurate of all). This part of the research is still pretty open and further
works are still going on in trying to figure out the correct functional. As for
concluding remarks we can have a look at figure 2.5 where the Jacob’s ladder
is given describing the order of accuaracy for different functionals in practice.
36
Figure 2.4: (a)Difference in energy per atom in the diamond phase and inβ tin phase of Si. The HSE functional agrees with the DMC results whilethe other functionals underestimate the energy difference (b)Formation en-ergy of the three lowest energy single interstitials in silicon (X, H, andT).Comparison to the DMC results demonstrate a steady improvement ofthe accuracy of the functionals as the order of density expansion increases,with quantitative agreement for the hybrid HSE functional.[41]
The development of approximations to the exchange-correlation functionals
over the past 20 years has improved the performance of DFT calculations,
and many scientists in the community think that progress up the Jacob’s
ladder will continue until energy differences can be determined to within 1
kcal/mol (“chemical accuracy”). Of course the numerical cost increases as
one climbs, and this may not necessarily bring more information.
2.5 Pseudopotentials
Pseudopotentials were originally introduced to simplify electronic structure
calculations by eliminating the need to include atomic core states and the
strong potentials responsible for binding them. Most physical and chemi-
cal properties of crystals depend to a very good approximation only on the
37
Figure 2.5: Jacob’s ladder of DFT schemes according to Perdew and collab-orators.
distribution of the valence electrons. The core electrons do not participate
in the chemical bond. They are strongly localized around the nucleus, and
their wave functions overlap only very little with the core electron wave func-
tions from neighboring atoms. Hence, the distribution of the core electrons
basically does not change when the atoms are placed in a different chemical
environment. It is thus justified to assume the core electrons to be “frozen”
and to keep the core electron distribution of the isolated atom in the crystal
environment. This is essentially a non-technical description of the otherwise
technical approach known as Frozen Core approximation. The first advan-
tage of the frozen–core approximation is that now less electrons have to be
treated and less eigenstates of the Kohn–Sham equations have to be calcu-
lated. Secondly the total energy scale is largely reduced when the core elec-
trons are removed from the calculation which makes the calculation of energy
differences between atomic configurations numerically much more stable and
tractable. Figure 2.6 is the schematic illustration of an atomic all–electron
wave function and the corresponding atomic pseudo wave function together
with the respective external Coulomb potential and pseudopotential. ΦPS
and VPS are the pseudo wave function and the pseudopotential, where as the
solid line is ΦAE, the actual wave function. In the scope of the present thesis
38
Figure 2.6: Schematic illustration of an atomic all–electron wave function(solid line) and the corresponding atomic pseudo wave function (dashed line)together with the respective external Coulomb potential and pseudopotential.
let me discuss some of the important classes of pseudopotentials. depending
on the properties they exhibit they can classified and used in various schemes.
2.5.1 Normconserving Pseudopotentials
The pseudopotentials we use today are constructed from ab initio calcula-
tions for isolated atoms. The Kohn-Sham equations for a single atom are to
be solved for a single atom of the chemical species for which we would like
to generate a pseudopotential. This can be done without much of a prob-
lem since due to the spherical symmetry of atoms the wave functions can
be written as a product of a radial function and a spherical harmonic. The
Schroedinger equation then reduces to one dimensional differential equations
for the radial functions which can be integrated numerically. Figure 2.6 rep-
resents such a typical result for a radial function from such an “all–electron”
atom calculation together with the corresponding external Coulomb poten-
tial. The aim now is to replace the effective all–electron potential within a
given sphere with radius Rcut by a much weaker new potential with a node
less ground state wave function to the same energy eigenvalue as the origi-
nal all–electron state which matches exactly the all–electron wave function
39
outside Rcut. An important aspect here is to understand why should this
happen at all? We can notice that the radial Schroedinger equation for a
constant potential and fixed energy E to the angular momentum l is a one-
dimensional ordinary linear second order differential equation which has two
linear independent solutions. But only one of the two solutions, Φl(r), is
regular for r → 0. We have to see that if we do not change the logarithmic
derivative
Ll(E) =d
drlnΦl(r;E)|Rcut
=Φ′
l(Rcut;E)
Φl(Rcut;E)(2.36)
while changing the potential inside the atomic sphere, then the wave func-
tions outside the sphere remain unchanged. For the energy of the eigen-
state of our all electron calculation EAEl the procedure is as explained in the
following section. The all electron wave function ΦAEl inside the sphere is
replaced by an arbitrary smooth nodeless function ΦPSl with the equivalent
logarithmic derivative at Rcut as the original all electron function. Since ΦPSl
is nodeless by construction the radial Schroedinger equation can be readily
inverted with this new function and with the eigenvalue EAEl of the all elec-
tron calculation to get the potential that has exactly the required property
for the more complicated case at hand. A lot of recipies are in publication
with which this can be implemented[44][45][46][47][48][49][51]. One further
important requirement is the essential normconserving condition. In sim-
ple words this means that the all electron and the pseudo wave function
inside the atomic sphere must have the same norm to guarantee that both
wave functions generate identical electron densities in the outside region. In
addition to this condition, the additional degrees of freedom in generating a
suitable pseudopotential can be employed to make the pseudo wave functions
as smooth as possible[52]. Evidently till now the above arguments claim the
existence of the logarithmic derivative of the effective all electron potential
only for the reference energy EAEl . Now however, if the chemical environment
of the atom in consideration is changed, evidently the eigenstates will be at a
slightly different energy. Hence of course, for a pseudopotential to be useful it
has to be able to reproduce the logarithmic derivative of the all electron po-
tential over a whole energy range. The transferability of the pseudopotential
40
to other chemical environments depends on this width of the energy range.
As reported in various articles, in particular the normconserving condition
guarantees such a transferability. And again, the pseudopotential should be
as soft (technical in sense) as possible. The term soft means that the number
of plane waves required to expand the pseudo wave functions should be as
small as possible. Both properties, transferability and softness, are closely
related to the cutoff radius Rcut, and compete with each other. Reports claim
that low cutoffs give pseudopotentials with a very good transferability[50].
However, increasing Rcut makes the pseudopotentials softer. Usually com-
promising balance between the two requirements is to be struck with. An
upper limit for Rcut, is given by half the distance to the next nearest atom in
the configuration for which one may want to apply the pseudopotential. If
this value is exceeded, we can see that there will not be any region between
the neighboring atoms left where we will recover the true all electron wave
functions. Hence, we can not expect anymore to get an accurate description
of the chemical bond between the two atoms.
2.5.2 Fully Nonlocal Pseudopotentials
From (2.36) it is evident that the logarithmic derivative depends on the
angular momentum l and hence a separate pseudopotential V PSl (r) for each
value of l needs to be constructed. This would call for the fact that, the full
pseudopotential for our atom therefore has to be a nonlocal operator. The
following equation describes the way in which it is done.
VPS = V PSloc (r) +
∑
l
V PSnl,l (r)Pl, V PS
nl,l (r) = V PSl (r)− V PS
loc (r). (2.37)
The pseudopotential V PSl (r) pertaining to one specific angular momentum
(usually the highest value of for which a pseudopotential has been generated)
is taken to be the local part of the pseudopotential. The nonlocal compo-
nents V PSnl,l (r) are defined as the difference between the original l dependent
V PSl (r) and this local part of the pseudopotential. As we know by now that all
V PSl (r) are identical beyond Rcut, hence we understand here that the nonlo-
41
cal components of the pseudopotential are strictly confined within Rcut. Pl is
a projection operator which picks out the lth angular momentum component
from the subsequent wave function. This construction guarantees that when
the full pseudopotential operator VPS is applied to a general wave function
each angular momentum component of the wave function experiences only its
corresponding part V PSl (r) of the potential. Since the projection operators
act only on the angular variables of the position vector r the pseudopoten-
tial is still a local operator with respect to the radius . The form (2.37)
is therefore called a semilocal pseudopotential. As reported, for numerical
efficiency however, it would be better to have the pseudopotential in a fully
nonlocal form as described in the article by Kleinman and Bylander[53]. In
the following section the technique as shown by Vanderbilt[54] to construct
a fully nonlocal pseudopotential, is discussed.
2.5.3 Vanderbilt Ultrasoft Pseudopotentials
Figure 2.7: Illustration of a strongly localized valence wave function insidethe atomic core region and the modified wave function in the Vanderbiltultrasoft pseudopotential scheme.
The general view is that it is very difficult to treat within a pseudopo-
tential scheme all elements with nodeless valence states (in particular those
with 2p and 3s valence electrons). For those elements the pseudo and the all
electron wave functions are almost identical. Since these valence electrons are
strongly localized in the ionic core region, many plane waves are required for
42
a good representation of their wave function which often makes calculations
for such elements highly expensive. In consequent works[54][55] Vanderbilt
has introduced a new type of pseudopotentials the so-called ultrasoft pseu-
dopotentials, in which the normconserving requirement has been relaxed.
Instead of representing the full valence wave function by plane waves, only a
small portion of the wave function is calculated within the Vanderbilt ultra-
soft pseudopotential scheme (as shown in figure 2.7). This allows to reduce
substantially the plane wave cutoff energy in the calculations. However there
is a price to pay in terms of complications in calculations. It is because now
the Fourier representation of the Kohn Sham equation becomes more compli-
cated. First, when the electron density is calculated the part of the electron
distribution has to be added back (which is represented by the difference be-
tween the solid and the dashed line in figure 2.7, the so-called augmentation
charges). Secondly, due to the relaxation of the normconserving condition,
the Bloch eigenstates will not be not orthonormal anymore. An overlap
matrix has to be introduced and the eigenvalue problem of the Fourier rep-
resentation of the Kohn Sham equations will transform into a generalized
eigenvalue equation. Next, the nonlocal part of the pseudopotential becomes
density–dependent now. Lastly ,due to these modification additional terms
in the force calculations have to be evaluated. However, the report claims
that the gain in computational cost by lowering the plane wave cutoff energy
outweigh in many cases the additional computational effort which is required
by these modifications.
With all those theoretical concepts discussed it is now time to talk about
the present day theoretical models with which people are trying to tackle
the probelm of charged defects in solids. It is also important to know the
experimental techniques involved in the study of point defects. With that
view we move on to the next chapter.
43
Chapitre 3
Dans ce chapitre, les differentes techniques utilisees pour traiter les
defauts charges dans les solides sont abordees en details. Je commence avec
les techniques experimentales deja utilisees dans le silicium depuis plusieurs
annees, et detaille les approches utilisees pour chaque type de defauts. En-
suite, j’aborde les differentes methodes theoriques. Les differences dans
le traitement de l’electrostatique sont rapportees et les avantages et in-
convenients des differentes methodes sont prises en consideration, montrant
la motivation de d’avoir une autre methode. Parmi les methodes decrites, on
trouvera celle de Makov-Payne, celle de Schultz, celle de Freysoldt ou celle
de Dabo.
44
Chapter 3
State of the Art: Experiments
and Theory
3.1 Introduction
The success of the Density Functional Theory (DFT) as a tool for ab ini-
tio calculation of various properties of solids has inspired scientists to use
it even to study defects (both charged and uncharged) in metals and semi-
conductors. In this chapter I wish to give a moderate review of the most
widely used and successful techniques in calculating the potential and con-
sequently the energies and forces of systems with charged defects. To model
a localised defect nested in the bulk host, the supercell approximation is
employed by most of the calculations. Reports of increasing usage of this
method to predict the structures of crytalline solids and even an extension
to the aperiodic systems of defects and molecules are found since a long time
now [60][61][62]. According to the available reports it is easier to deal with
the uncharged point defects in metals and semiconductors with this method
than to deal with defects in ionic insulators. In reality all the approximations
that goes into the calculations viz. the used pseudopotentials, the choice of
exchange-correlation functionals or the shape and size of the supercell, can
account for a difference in the predicted physical properties. Among the few
important causes for this trouble at hand are the defect defect interaction
45
due to localised charge and the self interaction of the strongly localised de-
fect states. Hence it is arguably evident that the formation energies thus
inferred through this periodic techniques may not find a strong ground in
terms of theoretical completeness and comparison with the experimental re-
sults. In this chaper I would discuss in brief the various modifications in the
calculations undertaken so as to find the correct way to calculate the forma-
tion energies of the charged defects in bulk systems. It is notworthy here to
mention that broadly we can tackle the problem at two different points in
the run of calculations. We can either start off with fixing the potential of
the system with physical arguments and consequently generate an accurate
density for the self-consistent loop in the DFT calculations or we can incor-
porate correction terms to the finally calculated total energies. It is worth
mentioning at this point that the total energies that are calculated from an
inaccurate and approximate potential are bound to deviate from the true
values (experimental in all physical cases). We will have an overview in this
chapter of all the formalisms and derivations of various correction factors.
3.2 Experimental Methods To Study Point
Defects
Before taking a plunge into the various theoretical methods of modern day
to handle the charged point defects it can be a interesting as well to know
few important experimental techniques that are used in abundance to study
point defects. The most popular set of experiments which where done to
study the defects in silicon were done by Corbett and Watkins[3][63]. Those
experiments were carried out with the methods known as Electron Param-
agnetic Resonance (EPR) or Electron Spin Resonance (ESR). The theory
of such experiments is based on the fact where the atoms of a solid exhibit
paramagnetism as a result of unpaired electron spins. Here transitions can be
induced between spin states by applying a magnetic field and then supplying
electromagnetic energy, usually in the microwave range of frequencies. The
resulting absorption spectra are described as electron spin resonance (ESR)
46
or electron paramagnetic resonance (EPR). ESR has also been used as an
investigative tool for the study of radicals formed in solid materials, since
the radicals typically produce an unpaired spin on the atom from which an
electron is removed. The study of the ESR spectra has been very effective
for the radicals produced as radiation damage from ionizing radiation. Study
of the radicals produced by such radiation gives information about the loca-
tions and mechanisms of radiation damage. The interaction of an external
Figure 3.1: Schematic diagram of the theory behind the ESR experiment.
magnetic field with an electron spin depends upon the magnetic moment as-
sociated with the spin, and the nature of an isolated electron spin is such that
two and only two orientations are possible. The application of the magnetic
field then provides a magnetic potential energy which splits the spin states
by an amount proportional to the magnetic field which is nothing but the
Zeeman effect. Then radio frequency radiation of the appropriate frequency
can cause a transition from one spin state to the other. The energy associ-
ated with the transition is expressed in terms of the applied magnetic field
B, the electron spin g-factor g, and the constant µB which is called the Bohr
magneton. In their work Corbett and Watkins studied the EPR spectrum of
the silicon vacancies and reported the geometry, energy of formation and sta-
bility of the different types of point defects (vacancies). Another technique
which they used was the Electron-Nuclear Double Resonance (ENDOR). EN-
DOR is a magnetic resonance spectroscopic technique for the determination
of hyperfine interactions between electrons and nuclear spins. There are two
principal techniques. In continuous-wave ENDOR the intensity of an electron
47
paramagnetic resonance signal, partially saturated with microwave power, is
measured as radiofrequency is applied. In pulsed ENDOR the radiofrequency
is applied as pulses and the EPR signal is detected as a spin-echo. In each
case an enhancement of the EPR signal is observed when the radiofrequency
is in resonance with the coupled nuclei. George Feher[64] introduced this
technique in 1956 so that interactions which are not accessible in the EPR
spectrum could be resolved.
A method to calculate the formation energies of the point defects is to directly
calculate the type and concentration of the defects. This was put forward
by the works of R. O. Simmons and R. W. Balluffi[65]. Here it is necessary
to measure the change of the lattice constant a, i.e. ∆a, and the change in
the specimen dimension, ∆l, (one dimension is sufficient) simultaneously as a
function of temperature. Which in other words is also called the differential
thermal expansion method. Here l is the length of the specimen which can be
thought as a function of the temperature and the defects i.e. l(T, defects).
The basic idea is that ∆l/l–∆a/a contains the regular thermal expansion
and the dimensional change from point defects, especially vacancies. This
is so because for every vacancy in the crystal an atom must be added at
the surface; the total volume of the vacancies must be compensated by an
approximately equal additional volume and therefore an additional ∆l. If we
subtract the regular thermal expansion, which is simply given by the change
in lattice parameter, whatever is left can only be caused by point defects.
The difference then gives directly the vacancy concentration. Some of the
other direct methods for measuring point defect properties can be classified
as the following.
• There can be measurements of the resistivity. This method is pretty
suitable to ionic crystals if the mechanism of conduction is ionic trans-
port via point defects. But one may never know for sure if what is
being measured is actually the intrinsic equilibrium because ”doping”
by impurities may have occurred.
• Measuring specific heat as a function of T is another method. While
there should be some dependence on the concentration of point defects,
48
it is experimentally very difficult to handle with the required accuracy.
• Measuring electronic noise can also be a direct method in this category.
This is a relatively new method which relies on very sophisticated noise
measurements. It is more suited for measuring diffusion properties, but
might be used for equilibrium conditions, too.
All the methods described above are for the equilibrium point defects. One
must undertake Quenching Experiments for the non-equilibrium point de-
fects. The basic idea behind these techniques is simple. If one has more
point defects than what one would have in thermal equilibrium, it should
be easier to detect them. There are several methods, the most important
one being quenching from high temperatures. The general methodology is
as follows.
• A wire of the material to be investigated is heated to some desired
(high) temperature T in liquid and superfluid He II (i.e. a liquid with
somewhat large heat conduction) to the desired temperature (by pass-
ing current through it). Surprisingly this is easily possible because the
He-vapor produced acts as a very efficient thermal shield and keeps the
liquid He from exploding because too much heat is transferred.
• After turning off the heating current, the specimen will cool extremely
fast to He II temperature (> 1K). There is not much time for the point
defects being present at the high temperature in thermal equilibrium to
disappear via diffusion; they are, for all practical purposes, frozen-in.
The frozen-in concentration can now be determined by e.g. measuring
the residual resistivity of the wire.
• The residual resistivity is simply the resistivity found around 0 K. It
is essentially dominated by defects because scattering of electrons at
phonons is negligible.
A relatively new way of looking at point defects is to use the scanning
tunneling microscope (STM) and to look at the atoms on the surface of the
sample. This idea is not new actually. Before the advent of the STM field ion
49
microscopy was used with the same intention, but experiments were (and are)
very difficult to do and severely limited. One idea is to investigate the surface
after fracturing the quenched sample in-situ under ultra-high vacuum (UHV)
conditions. This would give the density of vacancies on the fracture plane
from which the bulk value could be deduced. Although in these experiments
the vacancies can be seen, but there are other problems to tackle with. The
image changes with time - the density of point defects goes up.
In a nutshell I have tried to give an overview of the experimental techniques
that are involved in detecting the various properties of defects in the solids.
From the next section we would concentrate only on the theoretical models
to study defects in general.
3.3 Theoretical Methods To Study Point De-
fects
3.3.1 Treatment of Makov and Payne
In their article G.Makov and M.C.Payne[66] explained in length the tech-
nique of using Periodic Boundary Conditions (PBC) for different cases of
practical theoretical interest. They have argued the validity of using PBC
for calculations in solids on grounds like (i) ease of implementation in the cal-
culation (ii) compatibility with plane-wave expansions and consequent simple
calculations of forces in molecular dynamics simulations. (iii) availability of
an unified scheme for both periodic and aperiodic systems. In the article
they have proposed a detailed description of the electrostatics of the system
dealt with. The need for the correct electrostatics is felt as the convergence of
the total energy, with the increase of the size of the supercell, is determined
by the longest-ranged forces. Those forces are indeed electrostatic in nature.
The elastic forces can also be long ranged in solids and they are also treated
in the similar fashion in this particular report. In principle this work is a
continuation of the works by Leeuw[67] et al. The idea is to calculate the
total energy as correctly as possible because all the other physical quantities
50
are derived from the energy via the variational principle. For aperiodic sys-
tems using PBC, one is only interested in the energy, E0 in the limit L → ∞,
where L is the linear dimension of the supercell. The energy calculated for a
finite supercell E(L) differs from E0, because of the spurious interactions of
the aperiodic charge density with its images in the neighbouring cells. The
first case to investigate is the case with Neutral molecular species.
Neutral molecular species
Here for simplicity it is considered that the entire molecular charge density
is contained in the supercell. In practice, the electron density decays ex-
ponentially away from the molecular species and the supercell need only be
large enough for the density at the surface of the cell to be in this regime.
As reported the energy is absolutely convergent for the isolated molecule
which has no dipole moment. A careful conjecture about the nature of the
asymptotic behaviour of the energy would show that the energy will be,
E(L) = E0 +O(L−5). (3.1)
For an isolated neutral molecule that has a dipole moment the scenario is of
course different as the order of summation of the electrostatic sum should be
chosen with a bit more care. For an infinite lattice as considered in this type
of formalism that sum is not well defined. Hence the choice of the summation
area should be more of convenience with respect to the extent of the dipole
moment. In this case of course one finds the presence of a dipole dependent
term. It is also to be noted that for an aperiodic system, the absence of
this term will not change the value of energy in the limit L → ∞, as the
extra term is O(L−3). The results (figure 3.2) of a calculation from their
article show that energy calculated without the dipole term converges slowly
in comparison to the one, done with the dipole term. Another interesting
observation can be mentioned here. In general the dipole moment is reported
to be ill defined in PBC[68]. In principle, in a periodic solid, all choices of
supercell geometry relative to the charge distribution should give the same
energy. The electrostatic energy functional as menthioned in [66] infers that
51
Figure 3.2: The energy calculated without the dipole term converges slowlyin comparison to the one, done with the dipole term
different choice of the supercell will give different values of the dipole moment.
This is the manifestation of the fact that an infinite periodic solid with a
dipolar basis will not have a well defined energy. The limit L → ∞ dictates
the choice of the necessary supercell geometry. The supercell then must
create the picture of the aperiodicity in the cell as it would be in the bulk.
This is indeed necessary for the dipole moment to be invariant. For the case
considered in [66] the shape of the supercell is chosen to be a cube as the
result implies that dipoles on a cubic lattice do not interact and hence the
convergence of the energy would be O(L−5).
Nonneutral molecular species
Systems dealing with charged impurities or vacancies in crystalline solids
would call for total energy calculations in a periodically repeated electrically
charged system. And of course the value of the total electrostatic energy
diverges. In this scenario what is considered is the original charged system
52
immersed in a Jellium background, so the net charge is zero[62]. Then again
one can fall back to the procedure undertaken previously for the neutral
molecular species. The interaction between the Jellium background and the
charged species will decrease with the cell size and hence the convergence
will be slow, having a form of a power law in L. Evidently the charge density
of the immersed system consists of the density of the charged species and
Jellium density. This density again can be thought of as a contribution of
two charge densities giving rise to self interactions and an interaction with
each other. If we consider a charge placed at a particular point inside the
supercell, the self interaction of that charge will give rise to the term more
well-known as the Madelung energy of a lattice of point charges immersed in
a neutralizing Jellium.
EMadelung = −q2α
2L, (3.2)
where α is the lattice dependent Madelung constant. The mutual interaction
of the charge densities will also give rise to a very important term. As the
Jellium density depends on the supercell size hence this energy does not
converge as the previous one. The form includes a quadrupole term signifying
the interaction between the neutral defects. At the end what we get are two
correction factors. The assymptotic result for the total electrostatic energy
of a charged species on a cubic lattice finds the form as
E = E0 −q2α
2L− 2πqQ
3L3+O(L−5), (3.3)
where Q is the quadrupole moment and E0 is the desired electrostatic energy
of the isolated species. Results with ionization potential of a Mg atom are
shown in Figure 3.3. With all the corrections applied, it is evident that the
convergence is faster.
Aperiodic system in condensed matter
For aperiodic system in condensed matter the electrostatic energy can be
considered to be the sum of three major interactions. (1) The periodic charge
density interacting with itself. This is independent of L. (2) Interaction
53
Figure 3.3: (a)The ionization potential of a Mg atom calculated in cubicsupercells of side L (triangles), The same after including Madelung correction(squares), The same after the entire correction factor is applied (circles) (b)Expanded section of (a)
between periodic and aperiodic charge densities. This is also independent
of L.(3) The L dependent aperiodic densities located in different supercells.
Here also the main motive is to chalk out the asymptotic dependence of the
aperiodic density and its multipoles on the supercell dimensions. The first
source of L dependence is the one already discussed above, the case of changes
in the charge distribution induced by the interactions of the aperiodic charge
with its images. The second reason for the L dependence is the dielectric
response of the periodic medium to the aperiodic density. This is treated
nonelectrostatically and a correction factor containing the dilectric constant
is incorporated in the spirit of the arguments of Leslie and Gillan. The term
contains a quadrupole moment which is an artifact of the aperiodic density
whereas the dielectric constant is coming from the periodic density. With
this correction the final form of the total electrostatic energy becomes
E = E0 −q2α
2Lǫ− 2πqQ
3L3ǫ+O(L−5). (3.4)
In this case though, Q is the second radial moment only of that part of the
aperiodic density that does not arise from the dielectric response or from the
the Jellium. It is worthy of mention here that both Q and ǫ are properties
of the aperiodic density and the periodic density respectively.
54
3.3.2 Treatment of Schultz
A careful analysis by Peter A. Schultz[69] deals with some of the problems in
the Makov Payne formulation and employs a different method to avoid them.
Schultz concludes in his analysis that the Jellium approach fails to represent
the topology of the local electrostatic potential accurately. In fact he men-
tions in his article that the inaccuracy is to such an extent that for typical
semiconductors with sufficient band-gap energies the results obtained in the
Jellium approach are delusive. It is reported that tails of the potentials gen-
erated by local electrostatic moments of the defect overlap cell boundaries,
corrupting the local potential of the “isolated” defect. For supercells sep-
arated by vacuum gaps, e.g. molecules or repeated slabs, it is possible to
avoid, rigorously, the error of interacting multipole moments. He proposes a
technique as a generalisation of the Local Moment Counter Charge (LMCC)
method [70] to calculate a more accurate electrostatic potential within the
supercell. In this article the error in the predicted potential of the Jellium
approach in comparison to his LMCC method is also reported. The main
problem is that the local potential incorporates an error with the potential
tails of the periodic image charges which the compensating Jellium charge
does not take into account. The results show that the error in the potential
for semiconductors like charged Si and Ge are in the order of their band-gaps
viz. 1.00 eV for Si 64 atom cubic supercell and 0.62 eV for Ge 216 atom
cubic supercell. Increasing the cell size to reduce the Jellium potential error
is considered impractical in this article. In Figure 3.4 the error in the com-
puted electrostatic potential in a supercell calculation using jellium as the
neutralizing agent has been shown. As reported for vacuum gap supercells,
the LMCC method[70] is an alternate solution to Poisson’s equation that ex-
actly removes the spurious effects on the potential (and energy) of multipole
moments in the supercell. According to this formulation of Schultz, a model
density composed of a Gaussian array of charges is constructed in conformity
to the moments of the local system. Hence the system’s total charge distri-
bution comprises two parts, one part from this model density, and another
remaining half, which is devoid of any moments. Hence the potential for this
55
Figure 3.4: Error in the computed electrostatic potential in a supercell cal-culation using jellium as the neutralizing agent. The exact potential is com-puted analytically for a Gaussian density distribution in a cube, and usingan FFT after neutralization by jellium. The difference is plotted along a lineconnecting the centers of opposite faces and going through the center of thecell (inset).
part can easily be calculated with sufficient accuracy employing PBC and
Fast Fourier Transformation (FFT). The model density is treated as isolated
and only then the potential is calculated. Evidently the total potential is
the sum of these two, viz. the PBC potential and the Local Moment(LM)
potential,
φ(r) = φ′
PBC + φLM (3.5)
The idea is to create an isolated defect or molecule and blending the bound-
ary conditions to achieve the solution to the Poisson’s equation. For charged
defects in Si the potential φLM is found to be discontinuous and this phe-
nomenon needs to be avoided. The qrlong range asymptotic potentials mis-
align at cell boundaries. Again if the LMCC charge cannot be fixed by sym-
metry, the multipole moments of the defect are to be exactly determined.
In reality these are ill defined in bulk. Lastly the interaction of the LMCC
charge with the bulk material is to be taken into account. The potential of
an isolated charged defect extends outside the local cell volume and interacts
with bulk, but the LMCC potential is truncated at the cell boundary. The
Wigner-Seitz cell geometry around the position of each LMCC charge takes
56
care of the discontinuity in the potential. The multipole moments of the
defect can be calculated in two halves, contributions coming from the sum of
the spherical(momentless) atom charge distributions and the cell-dependent
moments of the perfect crystal reference and the defects. The last part of this
technique involves the interaction of the local charge with the bulk medium.
Schultz shows two Madelung integrals accounting respectively for (i)the in-
teraction between the LMCC and the potential from the crystal density and
(ii) the interaction between the LMCC potential and the crystal density in
WS volume.
EM =
∫nLM(r)δφref (r)−
∫
WS
φLM(r)δρref (r). (3.6)
The first term is a Madelung integral over all space between the LMCC
and the potential δφref (r) from the crystal density δρref (r) The second term
subtracts a local Madelung integral of the LMCC potential and a crystal
density in the WS volume. The energy of the dielectric response of the bulk
is calculated in same spirit as in the approach of Leslie and Gillan[71]. The
claim in this method is that it eliminates the spurious interaction of charges
between supercells and helps to compute the properties of charged defects
in bulk systems with the help of correctly computed electrostatic potential.
Results from three series of calculations are presented in the paper (figure
3.5): a 1D single strand of alternating Na and Cl, a 2D square single sheet
of NaCl, and 3D bulk NaCl. Also from Figure 3.4 it is to be noted that
the resulting energy error in the jellium-based calculation is not linear in
the charge. Doubling the charge doubles the error in the computed poten-
tial, and the energy is the integral of these two. Hence, the error in total
energy calculations scales at least as q2 , even neglecting the consequences
of screening. If results for singly charged systems bear uncertainties on the
order of a band-gap energy, results for multiply charged defects will be much
worse. In this study though the first principle calculations of total energetics
of charged vacancies in NaCl that accurately embody the correct electrostatic
interactions appropriate to isolated defects is reported with some criticism
of the Makov Payne formalism. It is demonstrated that the standard recipe
57
Figure 3.5: (a) The computed IP for a chlorine vacancy in a 1D NaCl chain,as a function of cell size. (b) The computed ionization potential for a chlorinevacancy in a 2D square NaCl sheet, as a function of cell size. Points frompolar NXN cells (inset: smaller 1X1 two atom cell) are plotted with squares,and nonpolarN
√2XN
√2 cells (inset: larger rotated four-atom
√2X
√2 cell)
with diamonds. (c) The computed ionization potential for a chlorine vacancyin bulk NaCl, as a function of cell size. Points from nonpolar cubic cells areplotted with squares, and points from polar fcc cells with diamonds. Thedashed line denotes the extrapolated asymptotic limit of the IP.
for studying charged defects in bulk systems, using jellium as a neutralizing
agent in supercell calculations, incorporates an error in the computed elec-
trostatic potential comparable to the energy scale of physical interest. The
generalized LMCC approach presented here claims to eliminate the unphys-
ical interaction of charges between supercells and enables the computation
of the properties of charged defects in bulk systems using the correct local
electrostatic potential.
58
3.3.3 Treatment of Freysoldt, Neugebauer and Van de
Walle
In their report Freysoldt[72] et. al. argues against the formalisms of the
correction factors of Makov and Payne and even the potential correcting
schemes Schultz. They argue that for realistic defects in condensed sys-
tems, however, the quadrupole moment correction term as stated in previous
works [66] do not always improve the convergence[73][74][75]. They have
shown that the Madelung correction greatly overshoots for small supercells.
Alternatively they have also argued against the technique of Schultz type for-
mulation where several scientists have suggested to truncate or compensate
the long-range tail of the bare Coulomb potential during the computation
of the electrostatic potential itself in order to remove the unwanted interac-
tions. According to Freysoldt[72] et. al. this scheme when applied to solids
suffers from ignoring the polarization outside the supercell. In their approach
they put forth a new idea which they claim (i) is based on a single super-
cell calculation, (ii) does not rely on fitted parameters, (iii) derives from an
exact expression applying well-defined approximations, and (iv) sheds light
on the problems encountered in previous schemes. The central idea of this
type of method revolves around the accurate calculation of the potentials
and proper treatment of the defect-defect interactions. The creation of the
charged defects is broken into three building steps which would account for
the final expression of the energy. First the charge for a single defect is in-
troduced by adding or removing electrons. At the same time all the other
electrons are frozen. This step results in an unscreened charge density. Ob-
viously enough the following step would incorporate all the other electrons
to screen this introduced charge. Lastly an artificial periodicity is imposed
upon along with the addition of a compensating homogeneous background
charge. With all these inputs and considering the unscreened defect charge
to be fully contained in a convenient zone of the supercell the screened lattice
energy of the charge with the compensating background is formulated out.
Here one must take care that when the electrons are allowed to screen the
introduced charge, the electron distribution forces an alteration in the elec-
59
trostatic potential Vdefect with respect to the neutral defect, which is given
by,
Vq/0(r) = Vdefectq − Vdefect0 . (3.7)
The resulting potential then is (up to an additive constant) a superposition
of the potentials Vq/0(r + R), where R is a lattice vector. Now this Vq/0 in
the reciprocal space would be easily formulated with the Fourier transforma-
tion. Hence the periodic defect potential is obtained from a Fourier series as
Vq/0(r). The interaction energy associated with the artificial potential due
to the periodic repetition of the charge comes out to be
Einter =1
2
∫
Ω
d3r[qd(r) + n][Vq/0(r)− Vq/0(r)], (3.8)
where n is the homogeneous background charge −q/Ω (Ω being the volume
of the supercell), qd(r) is the unscreened charge density and Vq/0(r)−Vq/0(r)
is the artificial potential due to the periodic repetition. The prefactor 12
accounts for double counting and the integral is restricted to the supercell.
Now the energy arising from the interaction of the background charge with
the defect inside the reference cell is calculated out in the following way,
Eintra =
∫
Ω
d3rnVq/0(r) = −q(1
Ω
∫
Ω
d3rVq/0(r)). (3.9)
Here n = −qΩ, the compensating homogeneous background charge. Adding
and rearranging (3.8) and (3.9) would allow the authors to explicitly calculate
the artificial interactions from parameters like ǫ, qd, Vq/0, which can then be
subtracted from the uncorrected formation energies obtained from the ab
initio supercell calculations. Theoretical value of ǫ is chosen (which can be
computed from various theories). For qd , it turns out that any reasonable
approximation to the defect charge distribution suffices since the sum of
lattice energy and alignment correction is not sensitive to the details of qd.
And Vq/0 is available directly from the DFT supercell calculations. The final
form of the formation energy for this paper is written down as (results of
unrelaxed Ga vacancy for a set of 2 × 2 × 2 supercells of the 8-atom cell are
60
reported here),
Ef (V qGa) = E(V q
Ga + bulk)− E(bulk) + E(Ga) + µGa − Elatq + q(EF +∆q/b),
(3.10)
where the last two terms the above equation are obtained by rearrangement
of the sum of (3.8) and (3.9) as
Einter + Eintra = Elat − q∆q/0. (3.11)
The Fermi energy EF is set equal to the valence-band maximum and the Ga
chemical potential µGa to that of Ga metal. The results are shown in figure
3.6. Using a point-charge model qd the ab initio corrected formation energies
Figure 3.6: Defect formation energies of the Ga vacancy in GaAs includingsupercell corrections as a function of the inverse number of atoms. TheFermi energy is set equal to the valence-band maximum and the Ga chemicalpotential to that of Ga metal.
of all calculations above 2 X 2 X 2 agree within 0.1 eV without any empirical
fit.
3.3.4 Treatment of Dabo et.al.
In the works of I. Dabo, B.Kozinsky, N.E.Singh-Miller, and N.Marzari[76]
argues in favour of their analysis to compute the correct electrostatic poten-
61
tial for charged systems in the framework of periodic calculations. The paper
even deals with extensive comparison with other methods of similar compu-
tation. In this work, they propose an alternative approach for correcting pe-
riodic image errors and show that energy convergence with respect to cell size
can be obtained at moderately available computational costs. The approach
advances by considering the exact electrostatic potential in open-boundary
conditions or any other boundary condition as the sum of the periodic so-
lution to the Poisson equation computed using FFT techniques (which are
inexpensive in reality) and a real space correction that can be obtained on
coarse grained meshes with multigrid techniques within a chosen degree of
accuracy. Hence basically the idea revolves around the trick of generating
a corrective potential on the basis of the boundary conditions. The correc-
tive potential is a mutual contribution from (in fact the difference between)
the potential arising from the Open Boundary Conditions (OBC) in absence
of an external electric field and the potential (satisfying the PBC) due to
the periodic translation of the same charge distribution. The corrective po-
tential satisfies much simpler, smoother differential equation (essentially the
Poisson’s equation).
∇2vcorr(r) = −4π(ρ), (3.12)
where vcorr is the corrective potential and ρ is an isolated charge distri-
bution satisfying the Poisson equation. It may be said here that a careful
and logical manipulation of the electron density ρ gives the corrective poten-
tial vcorr. The adjoining figure 3.7 gives an idea of the potentials for carbon
monoxide absorbed in on a neutral platinum slab and carbon monoxide ad-
sorbed on a charged platinum slab. The corrective potential seems to vary
more smoothly than the potentials following OBC or PBC. This scheme is
also tested with a point charge in a periodically repeated cubic cell and it is
found that the corrective potential needs to be expanded in a power series
exploiting the symmetry of the cubic cell for the convenience of computa-
tion. In Figure 3.8 we see the corrective potential in the presence of a point
charge q = +e in a periodically repeated cell of length L. The parabolic
expansion (valid up to third order) confirms that the point charge correction
62
Figure 3.7: Open-boundary electrostatic potential v, periodic potential v′
and electrostatic-potential correction vcorr averaged in the xy plane parallelto the surface for (a) carbon monoxide adsorbed on a neutral platinum slaband (b) carbon monoxide adsorbed on a charged platinum slab.
Figure 3.8: Corrective potential vcorr for a cubic lattice of point charges andits parabolic approximation in the vicinity of the origin.
or in reality the corrective potential is almost quadratic around the origin of
the system. Strikingly mathematical analyses yield that it is even quadratic
up to the third order for non-cubic lattices. Final form of the electrostatic
63
correction vcorr for an arbitrary distribution ρ would look like,
vcorr(r) =
∫vcorr0 (r − r′)ρ(r′)dr′, (3.13)
where the corrective potential generated by the uniform jellium and the sur-
rounding point charges is denoted by vcorr0 . The authors have compared
this scheme with some of the existing schemes and found that the correc-
tive potential gives out the Madelung constant and even a term similar to
the Makov Payne formulation, but differing slightly in magnitude though.
Three different cases are compared in this approach. One with a Parabolic
Point-Counter charge (PCC) potential. This one claims that this scheme
can correct periodic image errors up to quadrupole moment order. To obtain
higher-order PCC corrections, one would need to determine more terms in the
expansion of the point-charge correction beyond the parabolic contributions.
Another popular choice is to implement the Gaussian Counter charge (GCC)
formalism where the corrective potential is Gaussian. In this approach the
corrective potential is claimed to easy in computation. The GCC method is
somewhat similar to methods applied by Blochl[80] and LMCC method of
Schultz. At this point the authors felt the need for another approach which
is centered on the idea of the direct difference between the open boundary
potential and the periodic counterpart. This exact corrective potential is
coined as the Density Counter charge (DCC) potential by the authors. The
DCC potential is obtained by evaluating the Coulomb integral defining v
at each grid point in the unit cell. The PCC, GCC, and DCC potentials
for a charged pyridazine cation in a cubic cell of length L = 15 bohr are
plotted in figure 3.9. The PCC and GCC corrections are computed up to
the dipole order. First, it should be noted that the maximal energy of the
PCC potential is slightly above its GCC counterpart, reflecting the fact that
the Madelung energy of an array of point charges immersed in a jellium is
higher than that of a jellium-neutralized array of Gaussian charges. In ad-
dition, the maximal DCC energy is found to be above the Madelung energy,
proving that the dipole PCC and GCC corrections tend to underestimate the
energy of the system. Moreover, the parabolic PCC potential is not as steep
64
as its GCC counterpart, suggesting that the energy underestimation will be
more significant for the GCC correction. Because of the cubic symmetry of
the cell, the PCC and GCC potentials display the same curvature in each
direction of space. In contrast, the curvature of the DCC potential is not
uniform due to the nonspherical nature of the molecular charge density. This
shape dependence suggests that the accuracy of the GCC correction could
be improved by optimizing the geometry of the Gaussian countercharges.
Figure 3.9: PCC, GCC, and DCC corrective potentials for a pyridazine cationC4H5N
+2 in a cubic cell of length L = 15 bohr. The corrective potentials are
plotted along the z axis perpendicular to the plane of the molecule. ThePCC and GCC corrections are calculated up to dipole order.
∆Ecorr =1
2
∫v(r)ρ(r)dr. (3.14)
Equation (3.12) gives the final form of the corrected energy pertaining to
the corrected potential.
In Figure 3.10 the total energies of the Pyridazine cation have been plotted
with cell size in the different corrective schemes. From the figure 3.10 it is
evident that with the DCC scheme, the energy converges even more rapidly,
reflecting the exponential disappearance of energy errors arising from the
charge density spilling across periodic cells. Further treatments to other sys-
tems in one and two dimensional periodicity are still being worked upon. This
65
Figure 3.10: Total energy as a function of cell size for a pyridazine cationwithout correction and corrected using the PCC, GCC, and DCC schemes.The PCC and GCC corrections are calculated up to quadrupole order. Theinset shows a pyridazine cation in a cell of size L = 15 bohr.
scheme in general has been reported to have improved upon the drawbacks
of the Makov Payne or Schultz type formalism.
3.4 Conclusion
All the above mentioned formalims argue in favour of the periodic calcu-
lations for the charged systems. The idea is to correct for the erronious
description of the electrostatics because of the periodicity of the supercell
approach. We wonder if this is the only way to compute the physical quan-
tities of interest for systems with charged defects. Each process comes with
certain advantages over the other one, yet the fact remains as truth that
electrostatics of the charged system is not correctly portrayed in the PBC
schemes. The convergence issues are also of prime importance. All the above
mentioned methodologies exploit the DFT methods to calculate the total
energies and potentials required for the respective studies. The places where
we wish to probe in more is the search for a Free Boundary Condition (FBC)
scheme (under the DFT framework) with the exact electrostatics of the sys-
tem. We wish to check the validity of the pseudopotential approach as well
with a comparative idea of all the exchange correlation functionals and the
66
corresponding pseudopotentials for our case study. Before taking the ride
through the new method of treating charged deefects I would talk through a
chapter where a norm-conserving family of pseudopotentials is tested for its
accuracy. To make it even more accurate non linear core correction has been
added and further tested. The next chapter would deal with such a topic.
67
Chapitre 4
Dans ce chapitre est aborde l’utilisation de la correction non-lineaire de cœur
avec des pseudo-potentiels HGH afin d’ameliorer la precision des calculs ef-
fectues avec pseudo-potentiels a norme conservee, justifiant ainsi l’utilisation
d’un formalisme plus simple compare a la methode des projecteurs aug-
mentes (PAW). La correction non-lineaire de cœur est une facon de corriger
la linearisation apportee par l’utilisation de pseudo-potentiels. Ce chapitre
montrera la validite d’une telle approche ainsi que les resultats obtenues sur
les energies d’atomisation.
68
Chapter 4
Hartwigsen-Goedeker-
Hutter(HGH) pseudopotentials
with Non Linear Core
Correction (NLCC)
4.1 Introduction
As been discussed in the previous chapters, the fact DFT has proved itself
to be an extremely efficient approach towards the description of ground state
properties of metals, semiconductors and insulators motivates us to use it for
our first stretch of calculations. In the formalism of DFT, we have seen the
importance of a approximation –the exchange correlation functional. Within
a given range of error bar for this approximation it is already established that
DFT is the most widely used method for the electronic structure calculation
in the fields of solid state physics and quatum chemistry, as reports have
have been abundant in the community that, in the framework of DFT hun-
dreds of atoms can be dealt with immense accuracy and efficiency. Another
important aspect of the calculations that one may do with DFT is the use of
pseudopotentials. The pseudopotentials are used to simplify the lives of re-
searchers as they simply the calculations with less computation cost and they
69
also gives us a way to avoid the Dirac equation. Other wise we would have to
use the Dirac equation to take take into account the relativistic effects of the
all electron picture. It is known from theoretical quantum mechanics that the
wavefunction ocillates near the core for any species because of the core elec-
trons strong interaction with nucleus. This makes the all electron potential
very complex to deal with in computational techniques. The pseudopotential
is an effective (unreal of course) potential constructed to replace the atomic
all-electron potential (Full-potential) such that core states are eliminated and
the valence electrons are described by nodeless pseudo-wavefunctions. In this
approach only the chemically active valence electrons are dealt with explic-
itly, while the core electrons are ’frozen’, being considered together with the
nuclei as rigid and non-polarizable. Norm-conserving pseudopotentials are
derived from an atomic reference state, requiring that the pseudo and all-
electron valence eigenstates have the same energies and amplitude (and thus
density) outside a chosen core cutoff radius rc. Pseudopotentials with larger
cutoff radius are said to be softer, that is more rapidly convergent, but at
the same time less transferable, that is less accurate to reproduce realistic
features in different environments. Norm-conserving pseudopotentials em-
phasizes on the condition that, outside the cutoff radius, the norm of each
pseudo-wavefunction be exactly same to its corresponding all-electron wave-
function. We have used Hartwigsen-Goedeker-Hutter HGH pseudopotential
for this part of our calculations. The main idea is to use the pseudopo-
tentials for accurate calculations. One can argue of course the validity of
such a job. The emphasis is on the complexity-less approach of the norm-
conserving pseudopotentials to produce as accurate results as that of the
Projector Augmented Wave[80] (PAW) approach or that of the all-electron
calculations. The success of our approach would ensure a method other than
the PAW method for accurate calculations of various physical quantities.
Here in this work another important concept is discussed, the Non Linear
Core Correction (NLCC). This is an attempt to correct the otherwise im-
plied linearization of the pseudopotential. So in a nutshell this chapter will
deal with the validity of the results of calculations done with the Hartwigsen-
Goedeker-Hutter(HGH) pseudopotentials with NLCC . Another important
70
point to be mentioned here is the fact that this pseudopotentials are imple-
mented with the PBE exchange-correlation functionals on the wavelet basis
sets. Along with the exchange-correlation functional, the representation of
the Kohn-Sham wave functions is a prime concern for the DFT calcula-
tions. Usage of system independent basis sets such as plane waves[80]-[82],
wavelets[43], or real space grids[83][59] a very high level of accuracy can be
achieved. Basis sets of these kinds can be expanded systematically to achieve
convergence according to our level of confidence. Among the systematic ba-
sis sets wavelets are reported to have all the properties that are needed for
an accurate description of the system at hand. Systematic basis sets like
these even allows us to implement correctly and more conveniently the hy-
brid functionals. As reported earlier[43], such a method is implemented in
a DFT code, viz. BigDFT. As mentioned earlier it can be further clarified
that the goal of this work is to show the results of the implementation of
PBE functionals on wavelet basis. For the test the G2-1[84] set of molecules
are taken and the atomization energies of these 55 molecules are calculated
within the framework of PBE functionals implemented in BigDFT. The re-
sults of this work is duly compared with the known experimental values and
also with values reported from the calculation from VASP and Gaussian03
DFT codes[85].
4.2 Theory
4.2.1 Non Linear Core Correction
The need for NLCC
In their report[87] Louie and his colleagues have put forward a method to
extend the pseudopotential study primarily for the magnetic systems. The
aim is to deal with problems concerning properties of magnetic materials,
spin-density waves, magnetic effects on surfaces, localized impurity states in
defects, etc., without much computational trouble. The method also takes
care in keeping the level of accuracy as it has been in the nonmagnetic cases.
Previous to their works, to take into account the magnetic effects into pseu-
71
dopotential calculations, in their article, Zunger[88] proposed constructing
separate ionic potentials for the spin-up and spin-down electrons. Essen-
tially they devised a spin-dependent pseudopotential approach. And in here
the ionic pseudopotential in the solid depends on the spin density of the va-
lence electrons. This spin density is obtained by interpolation between the
ionic potential for the paramagnetic atom and that of the fully spinpolarized
atom.
The work of Louie and his collaborators differs from the work of Zunger with
a view that it is unnecessary and often undesirable to employ these spin-
dependent ionic pseudopotentials. As it is well known from the literature
mentioned in the previous chapters the exchange and correlation energy is
usually approximated by some local (nonlinear) function of the charge den-
sity, and the kinetic energy is found from the gradient of the now obtainable
singleparticle wave functions. As we have also seen that the charge density,
in the pseudopotential formalism, is divided into core and valence contri-
butions, and the energy of the core is assumed to be constant. Since the
energy of the core is not varying it is hence subtracted out from the total
density. Also it is to be kept in mind that quite often the core contribution is
absolutely neglected, and the total energy is given by the DFT energy func-
tional with the total charge density replaced by a (pseudo) valence charge
density. The electrostatic potential due to the interaction of the ions with
the core electrons is replaced by the pseudopotential. It is now pretty clear
that with this trick all the interaction between the core and valence electrons
is thus transferred to the pseudopotential. From the above arguments we can
understand that neglecting the core electrons poses a problem. This mode
of treatment linearizes the interaction which in reality is an approximation
to the kinetic energy and as well to the explicitly nonlinear exchange and
correlation energy. More explicitely, it is assumed in this fashion.
EXC(ρc + ρv) = EXC(ρc) + EXC(ρv), (4.1)
where EXC(ρc + ρv) is the total exchange correlation energy (non linear in
reality). The form shows how it is linearized. At this point it is also im-
72
portant to look for the cases where this approximation might have worked
very nicely. There can be cases where the core and the valence charge den-
sities are well separated in space. For example we can consider the cases
of various nonmagnetic molecules. Quite naturally there will be no serious
errors introduced here. Most of the results reported till date hold good only
because of this apparent spatial separation of the core and valence charge
densities. But of cource where there is significant overlap between the two
densities, the linearization, in particular of the exchange and correlation, will
lead to the problem of transferability of the pseudopotential. The fact that
the transferability is reduced, in turn, would incorporate systematic errors
in the calculated total energy. To deal with the problem one must under-
stand the basis nature and the origin of it. In the spin-density formalism the
exchange and correlation energy depends on the local spin density as well as
on the charge density. Now this additional spin dependence introduces the
additional nonlinearity which is the main point of discussion in this mod-
ule. Moreover it is the errors introduced by the linearization described above
that have caused the discrepancies between the experimental and theoretical
calculations and hence it is now necessary to use spin-dependent ionic pseu-
dopotentials. The scheme that is to be used now will formally be addressed
as Non Linear Core Correction (NLCC). NLCC treats these nonlinear terms
explicitly, and the need for separate spin-up and spin-down ionic pseudopo-
tentials is thus eliminated. What is better, the approach leads to significant
improvement in the transferability of the potentials. It is henceforth under-
standable that this will yield more accurate results both for magnetic and
non-magnetic systems.
Overview of the theory
Louie et. al., in their report[87] have given an example with the normcon-
serving pseudopotential scheme of Hamann, Schliiter, and Chiang[89](HSC).
To start with let me elaborate on the constraints with which the screened
atomic pseudopotentials V lion are constructed. It is to be kept in mind here
that these pseudopotentials are angular-momentum-dependent. The con-
73
straints are -
• The valence eigenvalues from the all-electron calculation and those from
the pseudopotential calculation agree for any configurational prototype.
• The all-electron wave functions and the pseudo-wave functions agree
beyond a chosen core radius, rc.
The two effective properties that are equally desirable and essential come
out from these basic constraints.
• The electrostatic potential produced outside rc, is identical for the all-
electron and the pseudocharge distribution.
• The scattering properties of the all-electron atoms are reproduced with
minimum error as the electronic eigenvalues move away from the pro-
totype atomic levels.
These two properties are claimed to be the reason for ensuring a reasonable
transferabihty of the pseudopotentials. The final bare-ion pseudopotentials,
V lion are extracted from the neutral potentials by subtracting from each neu-
tral V l the Coulomb and exchange and correlation potentials due to the
pseudovalence charge density, ρv(r). As for example, for a given angular
momentum component l and spin component σ, the ionic potential is given
by
V lσion(r) = V lσ(r)− Vee[ρ
v(r)]− Vxc[ρv(r), ξv(r)], (4.2)
where
ξv(r) =ρv+(r)− ρv
−(r)
ρv(r)(4.3)
is the spin polarization of the valence charge with the + and − signs denoting
the spin-up and spin-down electrons respectively. Ionic potentials generated
this way for the nonmagnetic case, i.e. for cases where ξ = O, have been
shown to be highly accurate in many applications[90]. In the above procedure
the basic assumptions are the frozen-core approximation and a decoupling of
the core charge in the determination of the exchange and correlation potential
seen by the valence electrons. At this point let us not forget an important
74
observation about the frozen-core assumption. It, however, implies a single
ionic pseudopotential which is independent of the spin polarization of the
valence states. Hence evidently it is the second approximation that gives rise
to the spin-dependent ionic potentials in previous work. In the HSC approach
the ionic pseudopotentials are dependent on the valence configuration, since
Vxc is a nonlinear function of the charge density and the valence charge does
not cancel in their formalism. It is explicitly written as,
V σxc(ρ
v + ρc, ξ) = [V σxc(ρ
v + ρc, ξ)− V σxc(ρ
v, ξv)] + V σxc(ρ
v, ξv), (4.4)
where
ξ(r) =ρv+(r)− ρv
−(r)
ρv(r) + ρc(r)(4.5)
It is actually clear that in the construction of the ionic potential, the terms
in brackets are included in the ionic potential as part of the core properties
which causes the problem in the transferability of the potential. This feature
is highly undesirable since it reduces the transferability of the potential.
In particular, for magnetic applications, the spin-density distribution of the
electrons can be extremely different both in magnitude and in profile as one
goes from the atomic case to the various condensed matter systems. Hence it
is improbable that any interpolation formula between spin-up and spin-down
potentials generated from atoms will work to a high level of accuracy and
satisfaction. To remove the dependence of the ionic pseudopotential on the
valence charge a simple and straightforward trick can be applied. The form
of the V lσion can be written as -
V lσion(r) = V lσ(r)− Vee[ρ
v(r)]− Vxc[(ρvr + ρcr), ξv(r)]. (4.6)
The total exchange and correlation potential, including the nonlinear core
valence term, is now subtracted out of the neutral potential. A smart and
effective implementation of this theory have been reported to have given
ionic potential highly transferable and essentially independent of the spin
polarization and the prototype atomic configuration. The following figure
with the results illustrate the acceptability of the above mentioned theoretical
75
method. In this regard the basic idea of implementation can be discussed.
Figure 4.1: Atomic term values and total energy differences between para-magnetic and fully spin-polarized atoms. Paramagnetic configurations are3s13p3 for Si and 4d55s1 for Mo. All energies are in eV. ∆E is the totalenergy difference between the paramagnetic and the spin-polarized configu-ration. Superscripts indicate the electron occupation and the + signs denotethe spin configuration for each orbital.
In the employment of the new potential, the core charge should be added
to the valence charge whenever the exchange and correlation potential or
energy is calculated. This core charge remains the same in all applications,
as we are considering things within the frozen core approximation. Hence, in
addition to the usual s, p,and d potentials, we need to retain the core charge
density which is computed once and for all in the same atomic calculations
as the pseudopotentials. It is certain that in an atomic calculation there is
no difficulty in representing the core charge. In a bulk calculation, however,
there are two practical considerations that must be taken care of. In any
pseudopotential calculation there are small, but unavoidable errors in the
calculated valence charge density. Usually this leads to a negligible error in
the total energy, but when the core charge is added, any inaccuracy in the
valence charge density inside the core region is multiplied by the core charge
and the error in the total energy will increase proportionally. To be sure of the
accuracy of the calculation, it is therefore absolutely necessary to treat the
effect of the core charge as a perturbation, and a change in the total energy
76
as a result of this correction should not be large. The above mentioned issues
can be handled by observing that the core charge has significant effect only
where the core and the valence charge densities are of similar magnitude.
One can therefore replace the full core charge density with a partial core
charge density which is equal to the true charge density outside some radius
rc and arbitrary inside. Louie[87] and his colleagues show that rc may be
chosen as the radius where the core charge density is from 1 to 2 times larger
than the valence charge density.
4.3 Tests for the pseudopotentials
The implementation of the PBE pseudopotential on the wavelet basis is
tested on two different levels. With a particular family of PBE pseudopoten-
tial, essentially the Hartwigsen, Goedecker, Hutter (HGH) normconserving
pseudopotentials, the first test were done. This family of pseudopotentials
are known to have the same limitations as that of the HSC pseudopotentials,
i.e. the core electron density is completely neglected for a prototype config-
uration. The results will show the performance of such pseudopotentials. It
can be mentioned here for the sake of information that the results were not
satisfactory and hence an improvement was sought for. The second level of
this test was to account for the limitations that we found to be inherent in
the method from the first test. The NLCC method was found theoretically
suitable for the cause and the next set of tests were done with this new family
of pseudopotentials corrected with the contribution from the core densities.
In the following sections the details of the calculations are explained with
results and analysis. The G2-1 test set has been taken and the atomization
energies for each of the molecules are calculated. The atomization energy
of a molecule E0(M) is defined as the diffrence between the energy of the
molecule ǫ0(M) and the sum of the energies of the constituent atoms ǫ0(X),
E0(M) =∑
species
nǫ0(X)− ǫ0(M). (4.7)
77
For all the atoms spin polarized calculations were done. It was taken strictly
into consideration that atoms were calculated in a symmetry broken ground
state. This state was found to be lower in energy than the spherically high-
symmetry solution. In the code care was taken as to satisfy the Hund’s rule
of electronic configuration for each constituent atom of each molecule. In the
following section let us take a look at a brief overview of Hund’s rule.
4.3.1 Hund’s rule
The rules formulated by Hund help to determine the total spin S, the total
spatial angular momentum L, and the total angular momentum J = L + S
of the ground state of an atom.
• If for a single atom there can be many electron states, the ground state
will be determined by the one having the largest total spin S.
• Among the states with the largest total spin, the ground state is the
one with the highest angular momentum quantum number L.
• For a shell less than being half-filled that ground state is the one with
the lowest value of J = L + S.
• For a shell more than being half-filled that ground state is the one with
the highest value of J = L + S.
These are the rules which are applied to the atoms for the calculations un-
dertaken in this work. But to apply these rules to the molecules care must
be taken so as to get the correct symmetry broken ground state. Hunds first
rules are reported to be good enough for most molecules. Most molecules of
interest have closed shell ground states, but there are a few exceptions. In
those cases Hund’s rule predicts that triplet lies below the singlet. In all the
calculations done in this work the Hund’s rules are applied with rigor and
care.
78
4.3.2 Calculational parameters
The main motivation for these calculations is to use the wavelet basis set and
to see the validity of the corresponding implementation of the NLCC within
the PBE pseudopotentials. In order to be highly accurate with atomization
energies for these calculations the parameters viz. hgrid and crmult for each
atom are tested for different values. Genovese et. al.[43] have mentioned the
theoretical convergence of the physical quantities calculated with respect to
these parameters. hgrid is the grid spacing and crmult is a parameter which
determines the extent of resolution zone that contains the exponentially de-
caying tails of the wave functions. Care has been taken to find the correct
combination of hgrid and crmult for each atom in the list of the G2-1 test
set of molecules. The hgrid and crmult are chosen in such a fashion that the
error in the total energy incurred in the calculation must not exceed 10−5
Hartrees, which is close to an error of .25 meV. Now from figure 4.2 it is
evident that for the atom hydrogen hgrid value of 0.30 and a crmult value
of 6 would give us results which are satisfactorily within our accuracy need.
Table 1 lists the values of the two parameters used for the elements in the
set.
Once the basic parameters to work with are acheived the main calculation
for all the atoms for the atomization energies are launched. As mentioned
earlier all the energies of the atoms are calculated in their spin polarized
state.
4.3.3 Atomization energy of molecules using PBE
The first test is performed with the PBE functionals in comparison with
the results of the plane wave methods as reported in the works of Paier[85]
et. al. From the figure 4.3 there are few interesting things that can be
understood. In a general statement one may consider the PBE functionals as
implemented on the wavelet basis set tends to underestimate the atomization
energy values as calculated from the PAW and G03 codes. Of course there
are atoms which have given excellent results in comparison to the all electron
calculations, but this is not the kind of test that will testify for the accuracy
79
hgrid in Bohr
crmult
1e-11
1e-10
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
err
or
in e
ne
rgy in
Ha
rtre
e
for Hydrogen atom
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
45
67
89
10
err
or
in e
ne
rgy in
Ha
rtre
e
0
0.0005
0.001
0.0015
0.002
0.0025
Figure 4.2: The error in total energy calculation of an H atom as a functionof the hgrid and crmult. The drawn dotted contour corresponds to the ∆Evalue of 10−7 Hartree.
of the impletation of PBE on the wavelet basis set. The main motive of this
work is to validate the implementation of the NLCC and its transferability.
The reasons for some atoms to be giving particularly good results can be
attributed to mainly two facts. One may believe as a primary reason that
the approximation to the densities, without taking into account the core
charges, were good enough for those prototypes. Again for molecules where
there are two or more atoms, respective errors in the calculation of the total
energies of each constituent atom cancelled the overall error. But obviously
in general the PBE functional as implemented on the wavelet basis set tends
to underestimate the atomization energy value. A careful observation of
the graph would help us to understand better the sources of errors. The
80
Table 4.1: The elements and the corresponding hgrids and crmults.
Elements hgrid crmultH 0.350 6Li 0.300 6Be 0.250 6C 0.250 6N 0.250 6O 0.200 6F 0.150 7Na 0.200 8Si 0.350 6P 0.350 7S 0.300 7Cl 0.250 7
molecules like O2, CO2, N2H4, N2, H2O2 have suffered the highest deviations
from the all electron calculations. The presence of oxygen in many of those
molecules hints towards the electronegativity of the atoms as well. Where as
the role of hydrogen may be excluded as it has only one electron and hence
there arises no question of the core charge. The nitrogen atom on the other
hand poses some significant problems. The reason to this may be because
of the strength of bonding of the nitrogen valence electrons. To associate
one nitrogen atom to any other atoms to form a molecure the core charge
density of the nitrogen might be perturbed and an inadequate representation
of that in the pseudopotential might result in the type of errors that are
exhibited in the figure. The following table shows the values as calculated
with BigDFT in comparison with reported values of VASP and G03 along
with the experimental values.
The average deviation of our results with this PBE pseudopotentials from
that of the G03 calculations is 5.197 kcal/mol and from that of Vasp is 5.015
kcal/mol. It is expected to improve upon the implementation of the non
linear core correction as with that formalism we hope to get closer to the
results obtained by all electron methods.
81
Table 4.2: Electronic atomization energies for the molecules in the G2-1 testset using the PBE functional. Energies are in kcal/mol.
Moleulcues AEPBE AEexpt ∆Expt− BigDFT% AEV asp AEG03 ∆V asp−BigDFT ∆G03−BigDFT
LiH 53.491 58.000 7.774 53.500 53.500 0.009 0.009BeH 55.660 48.000 15.958 55.500 55.600 -0.160 -0.060CH 82.819 84.000 1.406 84.700 84.800 1.881 1.981OH 105.566 107.000 1.340 109.700 110.100 4.134 4.534O2 130.381 118.000 10.492 143.300 144.000 12.919 13.619F2 44.641 38.000 17.476 52.600 53.000 7.959 8.359Li2 19.935 26.000 23.327 19.900 20.100 -0.035 0.165CO 254.750 261.000 2.395 268.600 269.100 13.850 14.350LiF 134.071 139.000 3.546 138.400 139.000 4.329 4.929SiO 190.156 191.000 0.442 195.600 196.600 5.444 6.444HF 136.815 142.000 3.651 141.500 142.200 4.685 5.385CN 185.236 179.000 3.484 197.500 197.700 12.264 12.464N2 226.867 227.000 0.059 243.700 243.900 16.833 17.033Cl2 64.757 57.000 13.609 65.800 65.800 1.043 1.043ClO 75.938 62.000 22.481 81.600 81.500 5.662 5.562HCl 105.664 107.000 1.249 106.300 106.500 0.636 0.836ClF 67.571 62.000 8.985 72.300 72.500 4.729 4.929Na2 17.497 19.000 7.911 17.700 18.100 0.203 0.603NO 157.290 153.000 2.804 172.000 172.500 14.710 15.210Si2 78.775 74.000 6.453 81.300 81.400 2.525 2.625NH 86.005 82.000 4.884 88.600 88.600 2.595 2.595P2 121.487 116.000 4.730 121.500 121.700 0.013 0.213S2 114.660 98.000 17.000 115.400 115.200 0.740 0.540SO 135.196 122.000 10.816 141.500 141.300 6.304 6.104CS 173.745 172.000 1.015 179.500 179.500 5.755 5.755CH3 306.418 306.000 0.137 309.700 310.100 3.282 3.682NH2 182.999 182.000 0.549 188.700 188.900 5.701 5.901H2O 225.689 233.000 3.138 233.700 234.500 8.011 8.811CO2 392.546 392.000 0.139 415.400 416.500 22.854 23.954SO2 269.089 253.000 6.359 281.100 280.700 12.011 11.611SH2 180.978 182.000 0.562 182.000 182.200 1.022 1.222H2O2 264.498 268.000 1.307 281.600 282.600 17.102 18.102HCN 311.872 313.000 0.360 326.300 326.500 14.428 14.628HCO 280.544 279.000 0.553 294.900 295.500 14.356 14.956HOCl 167.033 165.000 1.232 175.200 175.700 8.167 8.667NH3 293.184 297.000 1.285 301.700 302.300 8.516 9.116PH2 155.144 153.000 1.401 154.500 154.600 -0.644 -0.544PH3 239.176 241.000 0.757 239.000 239.300 -0.176 0.124NaCl 93.077 99.000 5.983 93.600 94.500 0.523 1.423CH4 413.802 420.000 1.476 419.600 420.200 5.798 6.398
SiH2(1A1) 147.577 154.000 4.171 147.900 148.000 0.323 0.423SiH3 222.992 226.000 1.331 222.200 222.600 -0.792 -0.392SiH4 313.536 324.000 3.230 313.300 313.700 -0.236 0.164C2H2 403.066 404.000 0.231 414.500 415.100 11.434 12.034C2H4 559.583 562.000 0.430 571.000 571.900 11.417 12.317C2H6 704.483 711.000 0.917 716.000 717.100 11.517 12.617H2CO 371.528 376.000 1.189 385.000 386.300 13.472 14.772N2H4 431.551 437.000 1.247 452.700 453.700 21.149 22.149
CH2(1A1) 174.512 182.000 4.114 178.800 179.100 4.288 4.588CH2(3B1) 193.571 189.000 2.419 194.400 194.600 0.829 1.029CH3OH 505.051 513.000 1.550 519.300 520.400 14.249 15.349CH3Cl 392.693 395.000 0.584 399.400 400.200 6.707 7.507CH3SH 470.621 473.000 0.503 477.800 478.600 7.179 7.979Si2H6 519.258 533.000 2.578 519.500 520.400 0.242 1.142
SiH2(3B1) 125.044 131.000 4.547 131.300 131.800 6.256 6.756
82
-20
-10
0
10
20
LiH
BeH CH
OH
O2
F2
Li2
CO
LiF
SiO HF
CN N2
Cl2
ClO
HC
lC
lFN
a2 NO
Si2
NH P2
S2
SO
CS
CH
3N
H2
H2O
CO
2S
O2
SH
2H
2O2
HC
NH
CO
HO
Cl
NH
3P
H2
PH
3N
aCl
CH
4S
iH2(
1A1)
SiH
3S
iH4
C2H
2C
2H4
C2H
6H
2CO
N2H
4C
H2(
1A1)
CH
2(3B
1)C
H3O
HC
H3C
lC
H3S
HS
i2H
6S
iH2(
3B1)
Diff
eren
ce in
the
atom
izat
ion
ener
gy fr
om th
e G
03 c
alcu
latio
ns
molecules of G21 set
pbewithoutnlcc-wavelet-basis-setpawG03
Figure 4.3: Electronic atomization energies for the molecules in the G2-1test set using the PBE functional. Energies are plotted in kcal/mol. The allelectron values from the Gaussian calculations are taken as the standard.
Geometries of the molecules
To complete the study with the PBE functionals the geometry of the molecules
were also investigated. A set of sixteen di-atomic molecules have been taken.
The molecules are made to go geometrical relaxation and then the bond
lengths are noted down. The following plot shows the variation of the bond
lengths with respect to the all electron calculations. The break criterion for
the geometry optimization in BigDFT was set to 10−5Hartree/Bohr which
is equivalent to 0.0005eV/A.
The table 4.3 and the figure 4.4 are an illustration of the fact that how
well the bondlength results have matched with the PAW and Gaussian cal-
culations. The bond lengths are already very good without the NLCC cal-
culations. One can although see a general trend. The bond length values
from all the three different theoretical calculations tend to overestimate the
83
Table 4.3: Tabular representation of theoretical and experimental bondlengths of 16 diatomics chosen from the reduced G2 data set. Bondlengthsare in angstroms.
Molecules Expt. PAW G03 BigDFT abs(∆BigDFT −G03)
in A in A in A in A in A×103
BeH 1.343 1.354 1.353 1.354 1CH 1.120 1.136 1.136 1.136 0Cl2 1.988 1.999 2.004 1.994 10ClF 1.628 1.648 1.650 1.645 5ClO 1.570 1.576 1.577 1.575 2CN 1.172 1.173 1.174 1.173 1CO 1.128 1.136 1.135 1.135 0F2 1.412 1.414 1.413 1.416 3FH 0.917 0.932 0.930 0.932 2HCl 1.275 1.287 1.288 1.286 2Li2 2.673 2.728 2.728 2.728 0LiF 1.564 1.583 1.575 1.575 0LiH 1.595 1.604 1.604 1.604 0N2 1.098 1.103 1.102 1.102 0O2 1.208 1.218 1.218 1.220 2Na2 3.079 3.087 3.076 3.070 6
84
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
BeH CH Cl2 ClF ClO CN CO F2 FH HClLi2 LiF LiH N2 O2 Na2
Diff
eren
ce o
f the
nlc
c b
ond-
legt
h va
lues
with
PA
W G
03 a
nd e
xpt.
List of selected diatomic molecules
pawg03pbe
expt refernce
Figure 4.4: Graphical representation of theoretical and experimental bondlengths of 16 diatomics chosen from the reduced G2 data set. Bondlengthdifferences are in angstroems.
experimental bond lengths. For most of the elements we find that the bond
lengths are very good accordance with the G03 calculations. As we can see
the problems for Paw with some of the molecules like LiF do not arise in
our case. The reason for the agreement for the bond lengths even with in-
accurate atomization energies can be explained as the following. Clearly for
the energies we see a problem with the reference energies of single atoms.
They are shifted in general from the accurate values. But since the bond
length calculations depend on the force calculations, which is essentially the
derivative of the energy surface. It appears that although the energy surface
is shifted by a value yet the shape of the surface is calculated with relatively
more accuracy and hence the force calculations are more accurate. Care
was taken by Paier[85] et. al. to minimise the dicrepancies. For example
in particular, the box size was increased, dipole correction was applied, and
85
the energy cutoff was increased. Yet the reason for this inaccuracy remains
open. It is also to be kept in mind here that for Li all electrons were treated
as valence electrons. There is another problem to be looked after both for
the PAW calculations as well as the wavelet calculations. The problem is
with the Na2 molecule. The Na2 molecule demand a tight convergence cri-
terion for the geometry relaxation because of the weak interaction between
the constituent Na atoms.
4.3.4 Atomization energy of molecules using PBE with
NLCC
Having seen the performance of PBE as implemented on the wavelet basis
(table 4.4 and figure 4.5) we are tempted to look for betterment in terms of
accuarcy towards the all electron calculation. With that view as discussed
in the precious sections the non linear core correction is the one interesting
way that could be thought of. A new family of pseudopotential is available
now under the PBE formalism with NLCC implemented in it. There are
parameters that can be played around with for respective elements to get
the most accurate and desirable results. Of course for atoms like hydrogen
and lithium these corrections are meaningless. Yet as it will be evident from
the following table and figures that considerable amount of improvement is
obtained with this implementation.
A single glance would show that the implementation of the NLCC in the
pseudopotentials have inproved the results by a large margin. Here also the
convergence parameters are kept the same as before for each prototype con-
figuration.
The significant improvement in the results can also be seen in the values of
mean absolute deviation of the atomization energies of BigDFT-Vasp and
BigDFT-G03. The average deviation of our results with this PBE pseu-
dopotentials from that of the G03 calculations is 1.03 kcal/mol and from
that of Vasp is 1.05 kcal/mol. Hence we can say at this point that in terms
of closeness to the calculations of the PAW and G03 the NLCC improves
the previous calculations by a factor of 5. Indeed this improvement is very
86
Table 4.4: Electronic atomization energies for the molecules in the G2-1test set using the PBE functional with NLCC correction. Energies are inkcal/mol.
Moleulcues AEPBE AENLCC AEexpt ∆Expt− BigDFT% AEV asp AEG03 ∆V asp−BigDFT ∆G03−BigDFT
LiH 53.491 53.489 58.000 7.778 53.500 53.500 0.011 0.011BeH 55.660 55.657 48.000 15.952 55.500 55.600 -0.157 -0.057CH 82.819 84.257 84.000 0.306 84.700 84.800 0.443 0.543OH 105.566 109.842 107.000 2.656 109.700 110.100 -0.142 0.258O2 130.381 144.184 118.000 22.190 143.300 144.000 -0.884 -0.184F2 44.641 54.495 38.000 43.409 52.600 53.000 -1.895 -1.495Li2 19.935 19.935 26.000 23.328 19.900 20.100 -0.035 0.165CO 254.750 270.133 261.000 3.499 268.600 269.100 -1.533 -1.033LiF 134.071 140.275 139.000 0.917 138.400 139.000 -1.875 -1.275SiO 190.156 200.394 191.000 4.918 195.600 196.600 -4.794 -3.794HF 136.815 142.809 142.000 0.570 141.500 142.200 -1.309 -0.609CN 185.236 197.722 179.000 10.459 197.500 197.700 -0.222 -0.022N2 226.867 243.491 227.000 7.265 243.700 243.900 0.209 0.409Cl2 64.757 66.572 57.000 16.793 65.800 65.800 -0.772 -0.772ClO 75.938 82.327 62.000 32.786 81.600 81.500 -0.727 -0.827HCl 105.664 107.164 107.000 0.153 106.300 106.500 -0.864 -0.664ClF 67.571 73.494 62.000 18.539 72.300 72.500 -1.194 -0.994Na2 17.497 17.481 19.000 7.995 17.700 18.100 0.219 0.619NO 157.290 171.982 153.000 12.407 172.000 172.500 0.018 0.518Si2 78.775 79.815 74.000 7.858 81.300 81.400 1.485 1.585NH 86.005 87.747 82.000 7.008 88.600 88.600 0.853 0.853P2 121.487 124.118 116.000 6.999 121.500 121.700 -2.618 -2.418S2 114.660 114.671 98.000 17.011 115.400 115.200 0.729 0.529SO 135.196 140.503 122.000 15.166 141.500 141.300 0.997 0.797CS 173.745 179.232 172.000 4.205 179.500 179.500 0.268 0.268CH3 306.418 309.409 306.000 1.114 309.700 310.100 0.291 0.691NH2 182.999 188.102 182.000 3.353 188.700 188.900 0.598 0.798H2O 225.689 235.435 233.000 1.045 233.700 234.500 -1.735 -0.935CO2 392.546 416.205 392.000 6.175 415.400 416.500 -0.805 0.295SO2 269.089 275.731 253.000 8.985 281.100 280.700 5.369 4.969SH2 180.978 181.665 182.000 0.184 182.000 182.200 0.335 0.535H2O2 264.498 283.128 268.000 5.645 281.600 282.600 -1.528 -0.528HCN 311.872 326.192 313.000 4.215 326.300 326.500 0.108 0.308HCO 280.544 293.879 279.000 5.333 294.900 295.500 1.021 1.621HOCl 167.033 176.017 165.000 6.677 175.200 175.700 -0.817 -0.317NH3 293.184 301.434 297.000 1.493 301.700 302.300 0.266 0.866PH2 155.144 155.808 153.000 1.836 154.500 154.600 -1.308 -1.208PH3 239.176 241.186 241.000 0.077 239.000 239.300 -2.186 -1.886NaCl 93.077 95.034 99.000 4.006 93.600 94.500 -1.434 -0.534CH4 413.802 419.374 420.000 0.149 419.600 420.200 0.226 0.826
SiH2(1A1) 147.577 148.582 154.000 3.518 147.900 148.000 -0.682 -0.582SiH3 222.992 223.630 226.000 1.048 222.200 222.600 -1.430 -1.030SiH4 313.536 315.582 324.000 2.598 313.300 313.700 -2.282 -1.882C2H2 403.066 414.388 404.000 2.571 414.500 415.100 0.112 0.712C2H4 559.583 570.842 562.000 1.573 571.000 571.900 0.158 1.058C2H6 704.483 715.702 711.000 0.661 716.000 717.100 0.298 1.398H2CO 371.528 387.062 376.000 2.942 385.000 386.300 -2.062 -0.762N2H4 431.551 448.464 437.000 2.623 452.700 453.700 4.236 5.236
CH2(1A1) 174.512 178.250 182.000 2.060 178.800 179.100 0.550 0.850CH2(3B1) 193.571 194.145 189.000 2.722 194.400 194.600 0.255 0.455CH3OH 505.051 519.241 513.000 1.217 519.300 520.400 0.059 1.159CH3Cl 392.693 399.551 395.000 1.152 399.400 400.200 -0.151 0.649CH3SH 470.621 476.638 473.000 0.769 477.800 478.600 1.162 1.962Si2H6 519.258 523.246 533.000 1.830 519.500 520.400 -3.746 -2.846
SiH2(3B1) 125.044 124.388 131.000 5.047 131.300 131.800 6.912 7.412
87
-20
-10
0
10
20
LiH
BeH CH
OH
O2
F2
Li2
CO
LiF
SiO HF
CN N2
Cl2
ClO
HC
lC
lFN
a2 NO
Si2
NH P2
S2
SO
CS
CH
3N
H2
H2O
CO
2S
O2
SH
2H
2O2
HC
NH
CO
HO
Cl
NH
3P
H2
PH
3N
aCl
CH
4S
iH2(
1A1)
SiH
3S
iH4
C2H
2C
2H4
C2H
6H
2CO
N2H
4C
H2(
1A1)
CH
2(3B
1)C
H3O
HC
H3C
lC
H3S
HS
i2H
6S
iH2(
3B1)
Diff
eren
ce in
the
atom
izat
ion
ener
gy fr
om th
e G
03 c
alcu
latio
ns
molecules of G21 set
without-nlccnlccpawG03
Figure 4.5: Electronic atomization energies for the molecules in the G2-1 test set using the PBE functional with NLCC. Energies are plotted inkcal/mol. The all electron values from the Gaussian calculations are takenas the standard.
significant.
Geometries of the molecules with NLCC
To complete the study with the PBE functionals with NLCC the geometry
of the molecules were also investigated in the same spirit as before. the same
set of sixteen diatomic molecules have been taken. The following plot in
figure 4.6 shows the improvement in the variation of the bond lengths with
respect to the all electron calculations. The break criterion for the geometry
optimization in BigDFT was set to 10−5Hartree/Bohr which is equivalent
to 0.0005eV/A which is same as the calculations without the NLCC.
88
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
BeH CH Cl2 ClF ClO CN CO F2 FH HClLi2 LiF LiH N2 O2 Na2
Diff
eren
ce o
f the
nlc
c b
ond-
legt
h va
lues
with
PA
W G
03 a
nd e
xpt.
List of selected diatomic molecules
pawg03
without-nlccnlcc
Figure 4.6: Graphical representation of theoretical and experimental bondlengths of 16 diatomics chosen from the reduced G2 data set with NLCC.Bondlength differences are in angstroms.
The bond lengths are observed to be more accurate as the figure and ta-
ble above suggest. It is also noteworthy here that in comparison to PAW
calculations our results have been found closer to that of the all electron
gaussian calculations for most of the molecules. This indicates the extent of
the success of the implementation of the NLCC in the pseudopotentials.
A closer analysis
Although as mentioned earlier the results have improved considerably with
the implementation of the NLCC yet a closer look into the graphical repre-
sentation (figure 4.7) would reveal some remnant discrepancies. The reason
89
Table 4.5: Tabular representation of theoretical and experimental bondlengths of 16 diatomics chosen from the reduced G2 data set with NLCC.Bondlengths are in angstroms.
Molecules Expt. PAW G03 BigDFT(NLCC)BeH 1.343 1.354 1.353 1.354CH 1.120 1.136 1.136 1.137Cl2 1.988 1.999 2.004 2.003ClF 1.628 1.648 1.650 1.649ClO 1.570 1.576 1.577 1.578CN 1.172 1.173 1.174 1.173CO 1.128 1.136 1.135 1.135F2 1.412 1.414 1.413 1.415FH 0.917 0.932 0.930 0.931HCl 1.275 1.287 1.288 1.287Li2 2.673 2.728 2.728 2.728LiF 1.564 1.583 1.575 1.575LiH 1.595 1.604 1.604 1.604N2 1.098 1.103 1.102 1.102O2 1.208 1.218 1.218 1.220Na2 3.079 3.087 3.076 3.070
90
-6
-4
-2
0
2
4
6
LiH
BeH CH
OH
O2
F2
Li2
CO
LiF
SiO HF
CN N2
Cl2
ClO
HC
lC
lFN
a2 NO
Si2
NH P2
S2
SO
CS
CH
3N
H2
H2O
CO
2S
O2
SH
2H
2O2
HC
NH
CO
HO
Cl
NH
3P
H2
PH
3N
aCl
CH
4S
iH2(
1A1)
SiH
3S
iH4
C2H
2C
2H4
C2H
6H
2CO
N2H
4C
H2(
1A1)
CH
2(3B
1)C
H3O
HC
H3C
lC
H3S
HS
i2H
6S
iH2(
3B1)
Diff
eren
ce in
the
atom
izat
ion
ener
gy fr
om th
e G
03 c
alcu
latio
ns
molecules of G21 set
nlccpawG03
Figure 4.7: Electronic atomization energies for the molecules in the G2-1 testset using the PBE functional with only NLCC correction . Energies are inkcal/mol. This is the zoomed version of figure 4.5.
for these discrepancies are to be sought after in detail.
In figure 4.8 what we get to see is the respective contribution of each el-
ement in a particular zone of absolute deviation from the G03 values. In
the figure Di-j is a zone of deviation in kcal/mol. Hence D0-1, D1-2, D2-4,
D4-7 are the deviation ranges of 0-1 kcal/mol, 1-2 kcal/mol, 2-4 kcal/mol
and 4-7 kcal/mol respectively. We can exclude hydrogen from our analysis
because of two main reasons. First of all, it has only one electron and hence
there is no question of fixing a special core density. Secondly it is the most
frequent element in the entire list. So it is quite evident that its contribution
is show in almost every range. The ranges which worry us the most are D2-4
and D4-7. Although we see some contribution of oxygen in those ranges,
91
0
5
10
15
20
25
30
35
40
45
Be C Cl F H Li N Na O P S Si Si
num
ber
of c
ontr
ibut
ions
to th
e co
rres
pond
ing
erro
r ra
nge
constituent elements
D0-1D1-2D2-4D4-7
Figure 4.8: Contribution of each element in a particular zone of absolutedeviation from the G03 values for the NLCC calculations.
a quick look at the table above would qualify oxygen as another of a fre-
quent element. Its association with elements like silicon and sulphur would
have forced its presence in those ranges of deviation. To this we come to
the elements like silicon, sulphur, phosphorous and nitrogen. The problem
with nitrogen is probably the same as discussed above. Its association with
hydrogen in N2H4 might have perturbed the core density which is still ill
represented in NLCC formalism. On the other hand the cases for silicon and
sulphur are not that straightforward. Silicon is semimetallic. An intuitive
approach would call for a smaller rc. The radius rc as described in the theory
section is an important parameter in this pseudopotential approach. Another
solution could be to play around with the amount of core-density to be used
in the NLCC part. At the end a general algorithm is what we have tried to
92
find that will be applied for all the elements in the same spirit to give more
accurate results. By this I mean we need to get rid of the deviation ranges
D2-4 and D4-7. The fact the silicon is indeed a bit tricky to handle, let
0
10
20
30
40
50
60
Be C Cl F H Li N Na O P S Si Si
num
ber
of c
ontr
ibut
ions
to th
e co
rres
pond
ing
erro
r ra
nge
constituent elements
D0-1D1-2
Figure 4.9: Contribution of each element in a particular zone of absolutedeviation from the G03 values for the PAW calculations.
us look in figure 4.9 where I have taken the data from the work of Paier[85]
and his collaborators. Here also we can see that silicon is found to be among
the very few to have contributed to the relatively higher deviation zone. In
this figure though we see contributions of carbon which in turn is not signifi-
cant with the NLCC calculations. Hence there is an indication regarding the
beahaviour of silicon in this particular picture. On the other hand in the
calculations with NLCC we have to look for other discrepancies arising from
sulphur and phosphorous. The problem with the phosphorous is indeed that
of a smaller rc. Calculations with smaller rc have yielded improved results
93
but the physics behind this improvement is still not better understood. The
question which is open at the moment is how to tackle the case of sulphur in
this general approach to create the pseudopotentials with NLCC.
4.3.5 Similar effort with Linear Spin Density Approx-
imation(LSDA)
Since the results with the calculations with the NLCC corrections to the
PBE pseudopotentials were found encouraging, efforts were made to do the
same under the LSDA formalism. The expected performance of those LSDA
pseudopotentials were on the same lines as we experienced with the PBE
+ NLCC pseudopotentials. In the article by Fournier[91] et. al. the cal-
-40
-30
-20
-10
0
10
20
30
CH
OH
O2
F2
Li2
CO
LiF
SiO CN N2
Cl2
ClO ClF
NO
Si2
NH P2
S2
SO
CS
CH
3N
H2
CO
2S
O2
H2O
2H
CN
HC
OH
OC
lP
H2
NaC
lC
H4
SiH
2(1A
1)S
iH2(
3B1)
SiH
3S
iH4
C2H
2C
2H4
H2C
ON
2H4
CH
2(1A
1)C
H2(
3B1)
CH
3OH
CH
3Cl
CH
3SH
Diff
eren
ce in
the
atom
izat
ion
ener
gy fr
om th
e LS
DA
cal
cula
tions
molecules of G1 set
abs-error lsda-BigDFt without nlccabs-error lsda-BigDFt with nlcc
Figure 4.10: Deviation of the BigDFT values from
culations of the atomization energies are being reported. They have calcu-
94
lated the atomization energies by different theories like Kohn–Sham density
functional theory (KS-DFT) with local spin density approximation (LSDA),
two KS-DFT gradient-corrected methods, one hybrid Hartree–Fock/KS-DFT
method similar to B3LYP, and the ab initio extrapolation procedures G1 and
G2. They have reported that the empirical corrections improved the LSDA
results greatly, while the other theories were improved slightly or not at all.
We have tried to compare our results with LSDA and NLCC with the results
of this report. In figure 4.10 the deviation of our calculations with their re-
ported values are plotted for both the cases. The two cases are viz. LSDA
without NLCC and LSDA with NLCC. as evident from the graph the results
are not encouraging. But the reason for such discrepancy can be thought
of as the error coming from the zero point energy which is neglected in the
cases of LSDA calculations.
4.4 Conclusion
In this part of the research we have tried to find an effective way to im-
plement the NLCC in the PBE and LSDA pseudopotentials on a wavelet
basis. The idea was to calculate the atomization energies to a ceratin de-
gree of accuracy and hence to use the pseudopotentials in further research.
Since wavelet basis sets have shown great promist in the DFT calculations the
need to implement various pseudopotentials on the wavelet basis set was felt.
As of now as we have shown that the results with the NLCC implemented
PBE pseudopotentials show good promise. Of course there is still room for
improvement. And also the need to understand the deeper physics for few
elements in the pseudopotential approach is felt. It must be highlighted that
the non-conserving pseudopotentials can be as precise as PAW if NLCC is
taken into acoount. We have also shown that the NLCC pseudopotential
seemingly improve the overall calculations by a factor of 5, which is indeed
a promising step in way of further research. Another important fact that
is revealed in the above analysis is the independence of the accuracy of the
geometrical structures from the jargon of NLCC in the pseudopotentials. By
this we mean that even with pseudopotentials where the entire core density
95
is considered to be constant the geometries of the molecules are correctly
predicted and understood. Having said that, we must also notice that im-
plementation of NLCC improved the geometry part as well. The need to
understand the LSDA formalism within the framework of NLCC is felt very
deeply. At present the question is open as to why the discrepancy of the
results would arise from the calculations of LSDA+NLCC as implemented
on the wavelet basis.
In the next chapter we will go into the main part of the thesis, the treatment
of the charged defects in silicon nanoclusters. Although those calculations
are done without NLCC as the work was simultaneous yet for future cal-
culations of charged defects one is tempted to use these NLCC corrected
pseudopotentials. In fact the next calculations without NLCC are sure to
give the geometries to a pretty high accuracy which is another reason to use
such NLCC uncorrected pseudopotentials.
96
Chapitre 5
Ce chapitre aborde le sujet des defauts charges dans le silicium. Jusqu’a
present, les defauts charges ont principalement ete etudies en conditions
periodiques (PBC). En effet, l’approche PBC est parfaite pour simuler un
solide infini. Mais, comme je le montre dans ce chapitre, elle apporte aussi
des artefacts. En particulier dans le traitement de l’electrostatique, ou des
traitements mathematiques important doivent etre utilises pour supprimer
les interactions non desirees entre repliques. Notre approche vise a simuler
correctement un defaut charge, tout en conservant de bonne proprietes pour
le materiau massif. Elle consiste a simuler le defaut dans un nano-agregat.
Le traitement de l’electrostatique est correct dans un nano-agregat et les
resultats obtenus peuvent etre extrapoles au materiau massif, comme il est
montre dans ce chapitre. Les perspectives de cette methode sont aussi
abordees ici.
97
Chapter 5
Charged Defects in Silicon
Nanoclusters
5.1 Introduction
The defects as it is mentioned in the introductory chapters are studied ex-
tensively under the formalism of DFT. It may be repeated as an emphasis
here as well that these calculations complement experiment in all possible
ways. They not only help to interpret experimental findings and link them
to an atomistic model of the relevant defects, but also provide additional
data such as formation energies, geometric structures, or the character of
the wavefunctions that cannot be obtained with state-of-the-art experimen-
tal methods. It is also worth mentioning again that the defect is usually
modeled in a supercell, consisting of the defect surrounded by many other
atoms of the host material, which is then repeated periodically throughout
space. The treatment is formally known as the Periodic Boundary Condi-
tion(PBC) approach. There have been reports[92] of the effectiveness of this
approach. However, it must be kept in mind that the use of supercells im-
plies that the isolated defect is replaced by a periodic array of defects. Such
a periodic array contains large imagineary defect concentrations, resulting in
artificial interactions between the defects that cannot be neglected by any
means. These interactions include overlap of the wavefunctions, elastic in-
98
teractions, and (in the case of charged defects) electrostatic interaction. A
large variety of approaches to control electrostatic artifacts exists in the lit-
erature, as has been reviewed recently by Nieminen[93]. As I have mentioned
in the earlier chapter (state-of-the-art chapter) supercell calculations for the
charged systems must always include a compensating background charge,
since the electrostatic energy of a system with a net charge in the unit cell
diverges[62][66]. The most common practice is to include a homogeneous
background, which is equivalent to setting the average electrostatic potential
to zero. By increasing the supercell lattice constant L, the isolated defect
limit can be realized in principle for L → ∞. Again it is to be remembered
here that the defect energy converges pretty slowly with respect to L (to be
precise it is L−1). The origin of this effect lies in the electrostatic interaction
of the defect with its periodic images which is unphysical. There is of course
an interaction of the charged defects with the compensating background. Its
magnitude can be estimated from the Madelung energy of an array of point-
charges with neutralizing background[62]. Makov and Payne[66] proved for
isolated ions that the quadrupole moment of the charge distribution gives rise
to a further term scaling like L−3. For realistic defects in condensed systems,
however, such corrections, scaled by the macroscopic dielectric constant ǫ to
account for screening, do not always improve the convergence[94][73]. As
mentioned in the chapter of the state-of-the-art there are other methods to
deal with the electrostatics of these charged defects amidst the image charges
and the background compensation in the PBC framework. All the calcula-
tion reported here are done with the DFT code BigDFT with the parameters
converged for silicon (discussed in the previous chapter). And all the calcu-
lations are spin average. The energy convergence criterion was kept to 10−4
Hartree and the force convergence criterion was kept to 10−5 Hartree/Bohr.
The total energy convergence criterion is based on the norm of the gradient
of the wave-function in the convergence cycle. In the following figure 5.1 and
figure 5.2 it shown in detail how the periodic extension of the supercell brings
in the unphysical elastic and electrostatic interactions between the real defect
and the imagineary defects. In this thesis a new approach is undertaken with
a view to study the charged defects under a simple yet correct electrostatic
99
Figure 5.1: A schematic representation of the image charges around the realdefect in PBC. The red dots inside the highlighted box are the real defects(charged or uncharged) and the rest are the product of the periodic extensionalong all directions.
model. The other method can be broadly classified as the cluster method.
In the cluster approach, a reasonably large cluster of atoms truncated from
the bulk with a defect at the center is used to simulate the bulk defect. For
this work we have created an isolated cluster of silicon atoms saturated by
hydrogen at the surface. Recent advances in electronic structure algorithms,
such as real-space methods[95][96] bring a new perspective to theoretical in-
vestigations of defects in materials in this cluster approach. Over the last few
years, these methods with the use of ab initio pseudopotentials have proven
their efficiency and accuracy for microscopic understanding of many physical
systems. In the works of Serdar Ogut and his collaborators reports have been
made on atomic and electronic structures of neutral and positively charged
monovacancies in bulk Si, investigated from first principles using a cluster
method[97]. In their work they have done calculations in real space on bulk-
terminated clusters containing up to 13 shells around the vacancy (about
200 Si atoms). Vacancy-induced atomic relaxations, Jahn-Teller distortions,
vacancy wave-function characters, and relaxation and reorientation energies
are calculated as a function of the cluster size and compared with available
100
Figure 5.2: A schematic representation of the elastic and electrostatic inter-actions between the real defect and the image defects PBC.
experimental data. But no reports have been made since then about the
applicability of this method in calculating the formation energy of vacancies
along with the formulation of a realistic charged-defect electrostatics. There
are two things which are very important in this approach. The first one is the
surface to bulk interaction as shown in figure 5.4. Since it is an attempt to
simulate bulk with a finite system we will definitely have the finite size effect
which can tracked down by the proper electrostatic analysis. This surface
electron density would essentially interact with the perturbed electron den-
sity around the vacancy to create a new phenomenon of trapped charges, as
will be shown in the following sections. This phenomenon of trapped charge
is the second aspect of this methodology. Again in our analytical model we
101
Figure 5.3: A schematic representation of the cluster method. The clustersare pacified by hydrogen atoms at the surface. The presence of only onearray of red dots inside the box ensures the simplicity and realness of thedefect representation.
would assume the clusters to be spherical, which in reality is true to a large
extent for really large clusters. But as shown in the schematic figure the clus-
ters are, in reality, polygons of some odd shape. Hence there is an element
of shape factor in this analysis as well.
5.2 Comparison of formulae (PBC vs. FBC)
5.2.1 PBC formalism to calculate the formation en-
ergy
Since in our case we are dealing with isolated clusters, hence the boundary
condition that we are using can be called Free Boundary Condition (FBC).
Let us now take a look at the respective expressions, under both the frame-
works of PBC and FBC, to calculate the formation energy of vacancies. In an
overview study of defect analysis Van de Walle[98] et. al. have summarised
a general formula to calculate the formation energy of the defects in PBC.
The formation energy of a defect or impurity X in charge state q is defined
102
Figure 5.4: A schematic representation of the surface to bulk interaction inthe isolated silicon clusters.
as
Ef [Xq] = Etot[Xq]− Etot[bulk]−
∑
i
niµi + q[EF + Ev +∆V ]. (5.1)
Etot[X] is the total energy derived from a supercell calculation with one impu-
rity or defectX in the cell, and Etot[bulk] is the total energy for the equivalent
supercell containing only the bulk species (for example Si, Ga, GaN etc.). ni
indicates the number of atoms of type i (host atoms or impurity atoms) that
have been added to (ni > 0) or removed from (ni < 0) the supercell when
the defect or impurity is created, and the µi are the corresponding chemical
potentials of these species. For the sake of clarification here we can describe
chemical potential as a physical quantity that represents the energy of the
reservoirs with which atoms are being exchanged. EF is the Fermi energy,
103
referenced to the valence-band maximum in the bulk. Due to the choice of
this reference, the users of this method need to explicitly put in the energy
of the bulk valence-band maximum, Ev , in the above mentioned expression
for the formation energy of charged states. A correction term DeltaV is
also needed to be added to the rest of the expression, to align the reference
potential in the defect supercell with that in the bulk. A problem when cal-
culating Ev is that in a supercell approach the defect or impurity strongly
affects the band structure. Therefore one cannot simply use Ev as calculated
in the defect supercell. To solve this problem a two-step procedure is used:
(a) The top of the valence band Ev is calculated in the bulk species by per-
forming a band-structure calculation at the Γ point and (b) an alignment
procedure is used in order to align the electrostatic potentials between the
defect supercell and the bulk. The fact that Ev found for the bulk, (i.e. in
a defect-free supercell) cannot be directly applied to the supercell with a de-
fect can be attributed to the long-range nature of the Coulomb potential and
the periodic boundary conditions inherent in the supercell approach. The
creation of the defect has been reported to give rise to a constant shift in the
potential, and this shift cannot be evaluated from supercell calculations alone
since no absolute reference exists for the electrostatic potential in periodic
structures. One common method is to align the electrostatic potentials by
inspecting the potential in the supercell far from the impurity and aligning
it with the electrostatic potential in the bulk species. This leads to a shift in
the reference level ∆V , which needs to be added to Ev in order to obtain a
more appropriate alignment.
5.2.2 FBC formalism to calculate the formation energy
In the scenario with FBC we try to build a formalism in similar terms. Here
also we calculate the total energies of the nano clusters with and without the
defects. To start with let’s put the formula as the following,
Ef = Etot[defect]− Etot[bulk] + µbulk−atom + qtrpd∆µ+ qsysµ. (5.2)
104
As we have done the calculations with vacancies in bulk silicon let me rewrite
the above equation in terms of bulk silicon for the sake of clarity.
Ef [Si(m− vacancies)] = E[n−m]Si+pH −EnSi+pH +mµSi + qtrpd∆µe + qsysµe,
(5.3)
where Ef [Si(m− vacancy)] is the formation energy of the vacancy (monova-
cancies or divacancies), E[n−m]Si+pH and EnSi+pH are the total energies of the
defect and the perfect system respectively. In the nano cluster there are in
general n Si atoms pacivated by p H atoms at the surface. m is the number
of vacancies created by removing atoms, (1 for mono vacancy and 2 for diva-
cancy). µSi is the chemical potential of bulk silicon which can also be thought
of as the energy of 1 bulk silicon atom. The most important and interesting
factors are the last two factors of the formula. Although the physical im-
portance of the quantity qtrpd will be explained in the following sections yet
it can be helpful to put in an overview of what it is in this context. qtrpd is
the amount of charged trapped near the defect as a consequential effect of
the defect formation. It is a section of the deformed electron density near
the defect and its value depends on the charge of the system. ∆ mue is the
difference of the Fermi energy between the defect and the perfect system due
to the occurence of the defect. qsys is the charge of the entire system for the
charged cases. It is of no use for the uncharged cases. µe is the Fermi level
of a particular charged system which can be considered as the reservoir from
where the charge is taken. To analyse the formula further we have to analyse
the electron density of the different defect systems.
5.3 Density analysis of the defect systems
5.3.1 Uncharged clusters
As it was stated earlier there will be surface effects in the approach within the
FBC formalism. The surface of the nanoclusters can cause some perturbation
in the electron density per unit volume near the center of the cluster. The
reason behind this can be attributed to the presence of the silicon-hydrogen
105
electr
on de
nsity
per u
nit vo
lume
electron density distribution of the perfect cluster
distance from the center in Bohr
surfacecenter
Figure 5.5: The variation of the electron density per unit volume aroundeach shell of the nano cluster (a) for all the clusters (b) for the cluster ofradius 1.2 nm
bonds at the surface which are electronically different from the bulk silicon-
silicon bonds. For our analysis we have studied clusters from 0.6 nm to
1.2 nm. The surface induced phenomenon may indeed bring about a charge
separation between the surface and the core center of the cluster. If we see at
the figure with the variation in the electron density per unit volume around
each shell of the nanocluster of a perfect cluster (without any defect) we
understand that the behavior near the center is similar to that of the PBC
calculations. As it is mentioned in the figure because of the surface and the
surface hydrogens the electron density trails off to zero. In the figure 5.5(a)
all the densities are plotted on the same plot so as to give a comparative idea
of the behavior of the densities in clusters of different sizes. We see from
the figure 5.6(b) that only near the vacancy (di and mono for this case) the
density per unit volume of the shell tends to differ from each other. For the
rest of it the curves are nicely superimposing each other. There is also a
question of the convergence of the observed and calculated physical quantities
with the cluster size. This question will precisely be address in the coming
sections. Let us have a look at the behavior of the density for the defect
states. In the figures 5.6 (a) and (b) it is shown how the electron density per
unit volume near the defect is disturbed for different cluster sizes. Near the
center we see because of the presence of the defects the electronic arrangement
106
Figure 5.6: The variation of the electron density per unit volume aroundeach shell of the nano cluster with a monovacancy (a) for all the clusters. (b)for all the clusters near the vacancy.
is different from that of the perfect clusters. A close look even at the density
distribution of the divacancy clusters would confirm the perturbation of the
density due to the presence of the vacancies. If we look carefully in figure
5.8 with the density plots of the largest cluster with a monovacancy and
divacancy respectively, we will see a clear difference in the electron density
distribution near the vacancies between the monovacancy clusters and the
divacancy clusters. It forces us to think that this might be the effect of
the surface and it triggers the thought of any charge separation between the
surface and the center of the nano cluster.
107
Figure 5.7: The variation of the electron density per unit volume aroundeach shell of the nano cluster with a divacancy (a) for all the clusters. (b)for all the clusters near the vacancy.
5.3.2 Charged clusters
In this section we will have a look around the vacancy in the charged clus-
ters. The main aim of the thesis is to create a formalism with which the
charged systems can be dealt without the image charge/defect interacts of
PBC formalism. Hence it is more interesting here to look into what happens
to the electron density distribution in the nano clusters when the defects
are charged. The overall density plots reveal a fact that near the vacancy
the electronic density is disrupted due to the presence of surface. And the
zoomed in figures also indicate that the charges of the system causes further
108
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0 5 10 15 20 25 30
elec
tron
ic d
ensi
ty p
er u
nit v
olum
e
distance from the vacancy in Bohr
electronic distribution of the neutral monovacancy cluster
surfacevacancy
surfacevacancy
1.2nm monovacancy1.2nm divacancy
dBulk
Figure 5.8: The difference in the electron density distribution near the va-cancies between the monovacancy clusters and the divacancy clusters. Thecluster size is 1.2nm
perturbation in the electron density near the vacancy. The question that
one may think about now is what is the effect of the system charge to the
electron density around the vacancy? We see from figures 5.8 , 5.9 and 5.10
of this section that only near the vacancy the density per unit volume of
the shell tends to differ from each other. For the rest of it the curves are
nicely superimposing each other. A further analysis with another tool would
eventually reveal the finer and clearer picture.
Mulliken analysis
Mulliken population analysis[99] is one tool which helps to estimate the par-
tial atomic charges from calculations carried out by few different methods of
109
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0 5 10 15 20 25 30
elec
tron
ic d
ensi
ty p
er u
nit v
olum
e
distance from the vacancy in Bohr
electronic distribution of the neutral monovacancy cluster
surfacevacancy
monodBulk1ncv2ncv1pcv
Figure 5.9: The difference in the electron density distribution near the va-cancies for a monovacancy cluster with different charge states.
computational chemistry, particularly those based on the linear combination
of atomic orbitals molecular orbital method. The charge or spin density dis-
tribution is studied as one of the most important properties of a molecule or
a cluster. Although there is no unique definition of how many electrons are
attached to an atom in a molecule or in a cluster it has nevertheless proven
useful in many cases to perform such population analyses. Due to its sim-
plicity the Mulliken population analysis has become a very familiar method
to count electrons to be associated with an atom. As an end result of this
calculation what is obtained is the gross charge in any atomic orbital. Which
is defined in this theory as the difference between the number of electrons
in an atomic orbital and the total number of electrons in the ground state
of the free neutral atom. For our research this analysis is done with the
110
Figure 5.10: The variation of the electron density per unit volume aroundeach shell of the nano cluster with a monovacancy (a) with 1 negative charge(b) with 2 negative charge (c) with 1 positive charge.
Figure 5.11: The variation of the electron density per unit volume aroundeach shell of the nano cluster with a monovacancy (a) with 1 negative charge(b) with 2 negative charge (c) with 1 positive charge , around the vacancy.
nano clusters of all sizes for all charge states and vacancies. The following
results will illustrate the difference in the population of the electrons of the
atoms around the vacancy. From figure 5.12 as one can understand that the
gross charge around each atom is pretty much the same for all the systems
except for the zone around the vacancy. Due to the nature of the bonds deep
inside the bulk we see that the number of electrons in a Si atomic orbital
is slighlty greater than the total number of electrons in the ground state of
the free neutral Si atom. As we go towards the surface due to the presence
of the hydrogens we experience the reverse the phenomenon. Right at the
surface the hydrogens tend to pull the electron density towards themselves
being more electronegative than silicon. Let us see for the fact that indeed
the discrepancy of the electron population is manifested near the vacancy or
at the center of the cluster. In figure 5.13 (a) we show all the gross charge
plots superimposed on each other, and we see that except for the zone near
111
Figure 5.12: The gross charge around the atoms in a cluster of size 1.2 nm for(a) perfect cluster (b) monovacancy cluster with charge -1 (c) monovacancycluster with charge +1 (d) monovacancy cluster with charge -2. The x-axisrepresents the atom ids in general. The hydrogens hence have the latter idsas the clusters are built atom centered.
the center of the cluster, in rest of the zones the plots are almost identical.
At this point it is felt necessary to comment on a small difference that is
observed for the case of the system with charge +1. As far as the hydrogens
are concerned they are more electronegative than Silicon and hence pull the
electron cloud towards themselves. But due to the presence of the positive
charge in the system and, as we will show how it also affects the charge
trapping near the vacancy, there is a Coulombic interaction between the pos-
itive charge of the system and the stretched electron cloud near the surface.
This gives rise to a variation in the values of the grosscharges for the system
with charge +1. This fact forces us to look deeper and closer in the zone of
the mono vacancy (figure 5.13 (b)). Here we see that corresponding to each
112
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0 100 200 300 400 500 600
Gro
ss
Ch
arg
ein
e
number of atoms
mono1ncv2ncv1pcv
perfect
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
0 5 10 15 20 25
Gro
ss
Ch
arg
ein
e
number of atoms
mono1ncv2ncv1pcv
perfect
Figure 5.13: The gross charge around the atoms in a cluster of size 1.2 nmfor (a) all clusters (b) all clusters around the vacancy.
charge state there is a different behavior of the grosscharge for that particular
charge state. Since the 15th atom or the third shell around the vacancy the
gross charges tend to behave similarly. From the above figure it is not easy
to quantify this effect. Now to quantify this effect we use the expression,
Qtrpd = Grosschargedefect −Grosschargeperfect, (5.4)
where Qtrpd can be thought of a residual charge which is trapped near the
vacancy, which is the difference between the grosscharge of the defect system
of any charge state to that of the perfect system. Now we can try to find out
whether this Qtrpd can be considered physical for a given analysis. Indeed if
we can see that this quantity is a converged one with the size of the nano
113
clusters then we can understand the significance of this physical quantity in
a broader physical sense. In figure 5.14 we see that with the cluster size
-0.25
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.7 0.75 0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2
Tra
pped
Charg
ein
e
mono1ncv1pcv2ncv
Radius of the nanocluster in nm
Figure 5.14: The convergenge of Qtrpd with cluster size.
the quantity Qtrpd converges to values for a cluster for each of the different
charge states. We can see from the figure that the values of the Qtrpd have
converged since the cluster size of 1.0nm, and it is only natural to consider
clusters from 1.0nm for further calculation of physical quantities. A striking
fact is that even if the value is very low yet the uncharged mono vacancy is
a bit negatively charged. And we can see that increase in the charge of the
system increases the value of Qtrpd. Hence Qtrpd is the most negative for the
most negatively charged system. This of course gives an impression of the
behavior of vacancies in a charged nano cluster. We can conclude from here
that indeed the defect in the nano clusters has a tendency to trap charges
and that quantity is dependent of the charge of the system.
114
Physical origin of the trapped charge
The presence of the trapped charges near the vacancy in the nano clusters
is a physical phenomenon which must have a relation any change in other
physical properties of the system. Results have shown as cited earlier that
presence of defects changes the HOMO-LUMO gap and the band structure
of any material, silicon for example in this thesis. Since we are dealing
with FBC we cannot do a band structure calculation, but a diagonalisation
method to calculate the HOMO-LUMO gap for the nano clusters can be used
instead. In this mode of calculation we calculate the HOMO-LUMO gaps
for the nano clusters of all sizes and charge states i.e. the difference between
the highest occupied molecular orbital(HOMO) and the lowest unoccupied
molecular orbital(LUMO). In the Free Boundary Conditions the HOMO-
LUMO gap is found to be heavily overestimated for smaller clusters. But
the HOMO-LUMO gap of the larger clusters are found to be converging
close to the bulk value. There are questions of quantum confinement in
the nano clusters of course. And the fact the HOMO-LUMO gap is heavily
overesmitated in very small nano clusters owe its origin to this confinement
effect. In figure 5.15 we have aligned HOMO levels so as to understand
the effect of the presence of the defects. The extreme-right column shows
the HOMO-LUMO gap as calculated from the 1.2 nm perfect cluster. The
middle column and the extreme left column are for the neutral monovacancy
and divacancy respectively. There we can see that indeed there are states
introduced within the HOMO-LUMO gap of the perfect nano clusters. Which
means creating one defect might mean otherwise to introduce energy states
within the HOMO-LUMO gap of a perfect system. We can investigate more
in the same spirit the system of the charged monovacancy. In figure 5.16 we
have shown a similar figure. Here we see that taking away one electron from
the system or giving one more electron to the system reduces the HOMO-
LUMO gap. In other words this can be though of as an effective way to
reduce the barrier of the electron flow in the bulk silicon. But there is
something which is found to be striking in this analysis. It seems that an
excess of negative charge in the nanoclusters increases the HOMO-LUMO gap
115
Figure 5.15: Representiing the change in the HOMO-LUMO gap due tothe presence of defects (uncharged) (-1 means mono vacancy and -2 meansdivacancy)
.
again. In reality the HOMO-LUMO gap of a neutral monovacancy cluster is
narrower than a monovacancy cluster of charge -2. We have already seen with
the mulliken analysis plot that increase in the negative charge of the system
will increase the negativity of the charge at the center. It may be because of
this charge an excess of charge is not welcome due to the coulombic repulsion
between the excess charge and the trapped charges. The argument is still an
open matter for further analysis and explanation. Now coming back to the
origin of the trapped charges again we must take into consideration the fact
that due to the presence of the defects and the charges there is a shift in the
chemical potential of the respective charge-defect systems. This shift can be
quantified as
∆µe = µeperfect − µedefect . (5.5)
116
Figure 5.16: Representiing the change in the HOMO-LUMO gap due to thepresence of defects (charged monovacancy with charges = -2,-1,0,1)
.
And the trapped charge Qtrpd is related to this shift in chemical potential
by the energy required for that charge to be brought at the center around
the vacancy in the cluster and we can quantify that energy as
Etrappedcharge = Qtrpd∆µe. (5.6)
This expression thus arrives in the formula of calculation of the formation
energy. Now the following motivation would be to look for an analytical
electrostatic model which must simulate this sort of a charge separation in
the nanoclusters.
5.4 Comparison of results obtained
Table 5.1 shows the performance of the model we have used to calculate the
formation energies. Experimental values in the spirit of the work by Watkins
117
Table 5.1: Table representing the comparison of the values of the formationenergies of monovacancies as calculated under PBC[75] and FBC(this work)and experiment[3].
Charge PBC FBC Expt.0 3.605 3.70 3.602-1 4.319 4.63 —-2 4.909 4.67 —1 3.545 3.86 —
and Corbett are still not available for the charged monovacancies and hence
it is still a point of time to talk about the accuracy of the model in discussion.
Yet we see our results are not far from the PBC values. Here some things
are to be kept in mind. The important thing is the fact that our motive
is to create a model without defect-defect interaction and devoid of any
background charge concept. In the following sections we will put forward an
analytical model to simulate the electrostatics of the given FBC formalism
which would indeed prove that our results have a very good chance to be
accurate with respect to the future experiments. The question of convergence
can be raised here and we will show in due course of this thesis that the
convergence issue can be tackled well with the electrostatic and elastic model
of the entire formalism.
5.5 Electrostatic model
In this work we have worked out an electrostatic scheme which appears fit
the atomistic model representation of the defects in bulk silicon and helps
to incorporate the concept of the trapped charges into an analytical model.
As the following figure will show in this approach we deal with an isolated
system free from all the interactions between the image charges and the real
bulk defect. Of course we have a surface to bulk interaction but as it will be
shown in the following analysis that this surface to bulk interaction indeed
helps to picturise the real electrostatic scenario of the cluster approach. In
the following section we will try to understand how this gives rise to the
118
Figure 5.17: Spherical shell of radius R with a smeared surface charge and anet charge at the center with the tiny sphere.
previously shown phenomenon of trapped charges.
5.5.1 Overview
Since the surface electron density interacts with the somewhat perturbed
electron density around the defect, it can be apprehended that this interac-
tion will cause some charge separation between the surface and the center
of the cluster. In this analysis we are dealing with the vacancies. Because
of the absence of the atoms there are dangling bonds from the neighboring
atoms of the vacancies. These electron density interacts with the surface and
as a consequence of the shift in chemical porential we find the occurence of
the trapped charges. The electrostatics of the entire system can be thought
of a problem where it is the objective to find out the elctrostatic energy that
can be stored in a cluster of ’n’ Silicon atoms passivated by ’p’ Hydrogen
atoms at the surface. The first model that is considered here is for that of
an uncharged cluster of radius ’r’. Although in reality the clusters are not
strictly spherical they are assumed to be perfect spheres in the model.
119
5.5.2 Details
Well as shown in figure 5.17 let the cluster be a spherical shell of radius R
with a net positive charge q placed at the center and a net positive chage −q
smeared all over the surface of the spherical shell. The spherical shell is filled
with a dielectric substance with a dielectric constant ǫr. Among the various
methods used to calculate the total electrostatic energy stored in any system,
the one that is used here is with the method of finding the electrostatic field
in different regions of the system. Then the electrostatic energy stored in the
system will be given by:
W =ǫ02
∫ǫrE
2 dτ (5.7)
where the symbols have usual meaning.
Let us divide the regions in our system into two different parts, viz. for the
electrostatic field inside the spherical shell, EI , and for the electrostatic field
outside the spherical shell, EII . If we consider a Gaussian surface outside
the spherical shell we can readily see that the total charge enclosed within
the Gaussian surface is 0 and hence the field outside the spherical shell, EII ,
is evidently 0.
Now to calculate the field inside the spherical shell we consider a sphericall
Gaussian surface along the surface of the spherical shell of our model. Here
the total charge enclosed is q. If we apply Gauss Law here we get:
∫EI . da =
Qenclosed
ǫ=
q
ǫ0ǫr(5.8)
By symmetry considerations we can see that the field inside is constant and
is radially outward. Hence we can easily deduce that the field inside the shell
is:
120
EI =q
4πǫ0ǫrr2(5.9)
Now to calculate the total energy stored in the system we fall back to equa-
tion (1) and integrate over the spherical volume element dτ = r2Sinθdθdφdr.
Hence the final expression of the total electrostatic energy stored in the sys-
tem is given by:
W =ǫ02
∫ǫr(
q
4πǫ0ǫrr2)2r2Sinθdθdφdr (5.10)
where the integrating limits for θ and φ are from 0 to π and 0 to 2π re-
spectively and hence integrating over θ and φ we get 4π. And the limit of
integration for r is from 0 to R. Hence from equation (4) we get:
W =q2
8πǫ0ǫr
∫ R
0
1
r2dr (5.11)
5.5.3 Modification
From equation (5.11) we can see that there is a problem, for the integrated
value at r = 0 is definitely blowing up. And it is evident from the fact that
the electrostatic energy of a point charge is indeed infinite. Hence we feel
the need of modifying our model a bit. Let us now assume another very tiny
spherical shell around the charge q. And we restrict ourselves in knowing the
electrostatic energy within the region between the two shells. Let the radius
of the tiny shell be R1. Now equation (5.11) becomes integrable again. The
total energy stored within the region of our interest is now given by:
W =q2
8πǫ0ǫr[1
R1
− 1
R] (5.12)
121
5.5.4 Remarks
The units of the quantities in the final expression of the total electrostatic
energy are all in S.I. units. Hence the energy that will be computed from this
expression will be in Joules(J). To compare with the total electronic energies
of our systems we need to convert this energy in to electron-Volts (eV). The
importance of this model will be explained in the following sections as we will
see the correspondence of the functional similarity of derived expression to
the actually calculated values. We wish to see whether this analytical model
is suitable to simulate the real electrostatics of the system. In the section
where we discuss the validity of this model we will show that the convergence
of the formation energy of the vacancies follow the same functional trajectory
as the electrostatics mentioned in this section.
5.6 Geometry of defects in nanoclusters
An important section of study regarding the defects is their geometry. A
defect essentially is an inconsistency in the geometrical structure whose pres-
ence will alter the electronic structure of the material. The change in the
electronic structure is often manifested in the geometrical changes that take
place due to the defects. The movement of atoms around a vacancy or a
foreign substitutional element can be thought of as a primary example. In a
report[100] D.V. Makhov and L.J. Lewis have talked about the distortions for
the divacancy in silicon. In another letter[101] S. Ogut and J. R. Chelikowsky
have reported similar distorsions around the divacancies in crystalline silicon.
The general understanding of these distorsions is based upon an electronic
phenomenon known as the Jahn-Teller distorsion.
5.6.1 Jahn-Teller Distorsion (JT)
The formal statement of the Jahn-Teller theorem states that in a nonlin-
ear molecule, if degenerate orbitals are asymmetrically occupied, a distortion
will occur to remove the degeneracy or in other words in an electronically
degenerate state, a nonlinear molecule undergoes distortion to remove the
122
degeneracy by lowering the symmetry and thus by lowering the energy. In
context of the study of defects it is found that due to the removal of an atom
to create a vacancy (or more) an untended degeneracy is enforced upon the
system. To remove that, the atoms around the vacancy would move in such a
fashion which would cause the total energy of the system to go down in value
to attain a more stable state. Hence the real manifestation of this electronic
effect of the removal of the degeneracy is found in the change in the geometry
of the system in and around the defect. Since the pioneering electron param-
agnetic resonance (EPR) studies of Watkins and Corbett[3], there have been
controversies about the electronic and atomic structures of the divacancy in
crystalline Si. These controversies have been concerned with the exact nature
of the above mentioned symmetry-lowering Jahn-Teller distortions that split
the degenerate deep levels. There are of course many reports which would
reveal the fact that the sense and the magnitude of these distortions, as in-
ferred from the EPR data and theoretical calculations, have been at variance.
There can be two different manifestations of these JT distorsion. The reso-
nant bond (RB) and the large pairing (LP) distorsions. In the study of D.V.
Makhov and L.J. Lewis[100] a schematic diagram (figure 5.18) shown the
difference between these two types. Although the results of Ogut and Che-
likowsky matched the experimental analysis of Watkins and Corbett there
were reports based on the results of ab initio calculations for divacancies by
Saito and Oshiyama[102] proposed the inverse distortion mode, the RB mode.
In the calculation as reported in works of D.V. Makhov and L.J. Lewis[100],
they found that the resonant bond distortion has slightly lower energy than
the pairing distortion ( 10 meV). And they also mentioned that the transi-
tion from one mode to the other occurs without any potential barrier and is
accompanied by a rotation of the symmetry plane of the divacancy. Hence
they claim that at room temperature, a divacancy should oscillate between
these two distortion modes.
123
Figure 5.18: Schematic view of LP and RB divacancy distortion modes..
5.6.2 Jahn-Teller Distorsion in the uncharged nanoclus-
ters
In their work[97] Ogut and his colleagues have studied the size dependence
of the relaxations of the atoms around the vacancy with respect to the shells
around the vacancy. Their findings were such which proved the inward move-
ment of the neighboring atoms towards the vacancy. They even found the
same for the charged cases. In the study that we undertook we also tried to
find whether it is possible to observe such distortions in these silicon nan-
oclusters. We have tried to investigate the geometry of the vacancies in
similar spirit. We did full geometry relaxation of the atoms and the cutoff
for the maximum value of force in all direction on each atom was held at
10−5Hartree/Bohr which is equivalent to 0.0005eV/A. In our analysis with
the neutral divacancies we found that depending on the size of the clusters
RB and LP distorsions take place around the vacancy. These distortions are
therefore dependent in a particular fashion on the cluster sizes and a general
statement relating the type of distorsion to that of the size of cluster can be
124
made. In the figure 5.19 we demonstrate the nearest neighbors of the neutral
divacancy of clusters 0.9 nm and 1.2 nm. In this figure as we see the nan-
Figure 5.19: JT distorsions and the distance between the nearest neighborsof the neutral divacancy for (a) 0.9 nm cluster and (b) 1.2 nm cluster
ocluster of radius 0.9 nm exhibits the nearest meighbours of the divacancy
in a LP distorsion yet the 1.2 nm cluster divacancy has undergone a RB
distorsion. The clusters of sizes more 0.9 nm upto 1.2 nm have all exhibitted
the same tendency. This fact can be addressed with a careful observation
to the surface effect of the nanoclusters. Although the nanoclusters are con-
sidered spherical in the electrostatic model yet they are not spherical and
they have the planes of the polygons on each face. The arrangement of the
planes are different in clusters of different size and hence the surface effect
is not homogeneous or consistent in nature for all the clusters. And we may
be forced to think that this effect in the nanoclusters owe its origin to the
arbitrariness of the arrangement of the surfaces. To understand the effect of
the surfaces let us have a look at the monovacancies, figure 5.20. Here also
we can see the JT effect around the monovacancy. If we look carefully we
will notice the movement of the nearest neighbors to the vacancy in the two
clusters appears different even if viewed from the same direction. By this
I mean the movement of the neighbors in a preferred direction. Given the
isotropy of an infinite bulk this indeed is not an issue. One can argue that
in a bulk (which we are trying to simulate through our nanoclusters) these
125
Figure 5.20: JT distorsions and the distance between the nearest neighborsof the neutral mono vacancy for (a) 1.1 nm cluster and (b) 1.2 nm cluster
two apparently different movements are equivalent. But for our system as it
is finite the movement of the atoms are of course influenced by the surfaces
present in a particular cluster. A schematic diagram (figure 5.21) will be
helpful to understand this fact. For the sake of completeness here it must be
Figure 5.21: The two modes of JT distorsions as observed in the nanoclus-ters. When viewed from a particular direction in a finite system they appeardifferent of course. The box inside is the vacancy and the round balls are theneighboring atoms.
mentioned that out of these two different movements both are equally likely
as it does not depend on the cluster size as it was for the divacancies.
126
Table 5.2: Table representing the extent of the deformations in the distancebetween the nearest neighbors of the monovacancy in different charge states.Avg.4 is the column displaying the average distance between the 4 neighborswhich are at same distance from the vacancy and Avg.2 is for the averagedistance between the 2 neighbors which are at the same distance from thevacancy.
Charge Avg.4 Avg.2 % Deformation0 3.572 2.996 -16.13-1 3.464 3.058 -11.72+1 3.8 3.556 -6.42-2 3.094 3.423 +10.63
5.6.3 Jahn-Teller Distorsion in the charged nanoclus-
ters
Let us now investigate the extent of this JT distorsion in the charged systems.
As we have already seen that the presence of charge in the system affects
the electron density near the vacancy it is quite natural to assume that the
effect will be seen in the geometry of the system as well. In fact, as it is
already mentioned before, the geometry is the manifestation of the electronic
changes in the system due to the presence of the defects. Let us have a
look at the comparative JT distorsions in the neutral and variously charged
monovacancy clusters.
Table 5.2 along with figure 5.22 exhibits a very strange fact. Indeed from
figure 5.22 we understand that the presence of charge in the system and
due to the effect of the trapped charges near the vacancy the geometrical
manifestation of the JT distortion is not the same in all the charged clusters.
In fact we see that for all the clusters there is a tendency of 2 atoms moving
such that they share a similar distance from the vacancy and their distances
from the other atoms are larger of course. This is also shown in the schematic
figure as mentioned before in context of the two types of distortions in the
finite clusters. And we can have a measure of the deformation from the above
table. The table shows that for charge states 0, -1 and +1 the distance
between the atoms grouping together is respectively 16.13, 11.72 and 6.42
127
Figure 5.22: JT distorsions and the distance between the nearest neighborsof the monovacancy for 1.2 nm cluster with charges (a) 0 (b)-1 (c)1 (d)-2
percent less than the atoms being apart. Where as for the charge state -2
the distance between the atoms grouping together is 10.63 percent more than
that of the other atoms which have moved apart. In this regard we may recall
that even from the HOMO-LUMO gap analysis in the previous section we
have known the system with -2 charge behaves differently from the others.
This different mode of JT distortion must therefore be caused due to the
effect of the trapped electron density near the center of the nanocluster.
128
5.7 A simple force model to simulate the JT
distorsion
Figure 5.23: Schematic representation of a simple force model of JT distorsion.
Owing to the fact that the atoms are moving as a manifestation of the
JT distortion a simple force model can be visualized in accordance with
the overall picture. In the schematic figure (figure 5.23) of the model the
inward movement of the atoms towards the center of the cluster (around the
vacancy) is represented. Due to the movement of the atoms the volume of
the defect is changed by dV . The forces exerted on this defect volume are
from all possible directions. And hence we can imagine a contour of circular
areas as shown in the figure. The force per unit area for each such circular
area is the pressure P exerted on the volume dV and the work done to create
dV would be given as the elastic energy,
Welastic = PdV. (5.13)
and
P =Force
Area=
C
r2(5.14)
where C is a constant and r is the radius of the nanocluster.
129
5.8 The combination of the electrostatic and
simple force model
If we fall back on the electrostatic energy stored in the cluster to bring all
the charges together the general form of the expression would look like,
Welectrostatic = A− B
r. (5.15)
Hence if we combine the two models, i.e. electrostatic model and the simple
force model of JT distorsion we will get the total energy required to form the
vacancies in a nanocluster with the JT distorsion. Therefore we can say that
the total energy required to form the vacancies in a nanocluster with the JT
distorsion must have the analytical form of,
Eformation = A− B
r+
C
r2.(5.16)
Now we can check the validity of this functional form with different values
of the radii of the nanoclusters and then can simultaneouly plot the values
of the calculated formation energies under the FBC formalism. This as will
be explained in the through following figures will also help us to understand
the convergence of the value of the formation energy in relation with the
cluster size. As it is evident from the fitted curve (figure 5.24) that the
value of the formation energy seems to be converging with the cluster size
and hence the values quoted from the calculations of the 1.2 nm clusters can
be considered to be accurate under the given formalism. This curve also
indicates the physically acceptable nature of the models discussed above.
The points which are away from the analytical curve are of course results of
the finite size effect as one must experience in calculations like these. It also
emphasies the presence of a strong surface effect when the clusters are really
small. Hence it can be argued that the physical significance of the formation
energy as calculated from the cluster of size 1.2 nm can be taken as standards
for future bulk calculations.
130
-1
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
form
atio
n en
ergy
in e
V
radius of nanoclusters in nm
A - B/r +C/r2calculated plot
Figure 5.24: The comparison of the calculated formation energies with theanalytical model and the convergence of the formation energy value withrespect to the cluster size
.
5.9 Conclusions
As concluding remarks the following observations can be considered as the
result and immediate finidings of the research.
• The electrostatics of the nanoclusters with defect is relatively simple
and easy as compared to that of the PBC systems. The requirement
of a background compensating charge is not required.
• It is also noteworthy that the defect-defect interaction along with in-
teraction of the image charges can be avoided for the study of charge
systems. Of course there is a surface to bulk effect which is an artefact
of the FBC model but it is relatively simpler to deal with in the con-
text of finding the formation energies of the defects and studying there
geometries.
131
• The geometry of the charged defects in nanoclusters manifests most the
effect of the surface. This is a bit far stretched one may presume. But
indeed the surface induces the occurrence of trapped charges near the
vacancy. And we have show that the geometry for each charged state
is different from one another and it is obviously due to the different
values of the trapped charges. Hence we see different deformations in
the JT representations for different charge states. Also the forms of
distortion like the RB and LP distortions are observed in this kind of
an approach.
• The phenomenon of trapped charges is observed as an important feature
of the FBC model which allowed us to understand the electrostatics of
the nanoclusters with or without defects. We may even emphasize the
ability of the vacancies to trap charges in a solid as we have found
and proved in our analyses. Even for the case of PBC we have seen if
we somehow manage to keep a charge in the system far enough from
the vacancy (one cannot actually be very far from the vacancy in PBC
because of the periodicity itself) it tends to trap that charge. Hence in
general one may infer that vacancies trap charges.
5.10 Future perspectives
5.10.1 Different other defects
This analysis can be extended to the study of other defects as well. In fact
even within this research we have tried to get similar results for the charged
di vacancies as well. But unfortunately the required criteria for an acceptable
value of forces on the atoms are not reached. The reasons may be hidden
in the behavior of trapped charges for the divacancies. The electron density
around the divacancy must be disturbed in a greater extent due to the surface
and a spin polarized calculation of the nanoclusters would be preferred to
look into the matter more deeply. All the calculations up to now are spin
average calculations. But to do a spin polarized calculation on a cluster of
132
1.2 nm in BigDFT, the expense is huge and hence it is a bit difficult to tackle
at the moment.
5.10.2 Migration energies
Another method to verify whether the results obtained by this method are
believable or not is to check for other quantities which are reported in experi-
ments. One such method is to calculate the migration energy of a vacancy. In
figure 5.26 we can see a migration of vacancy pictorially. In reality we have
Intermediate state
Energy we need
Figure 5.25: Scematic diagram to explain the concept of the migration energy.
tried already to calculate the migration energy of monovacancy within the
framework of our model. We have used the DIIS[103] method to find the mid-
dle point of the journey of the vacancy from one place to the other. As per
our calculations we have found out that the migration energy of the neutral
monovacancy in 0.5 eV. As reported in the report by Watkins[63] the experi-
mental value of the migration energy of the neutral monovacancy is 0.45 eV.
This result encourages us to continue to do the same for other charge vacan-
cies. Indeed along with these calculations another form of calculation is still
under process. To get the contour of the entire path of migration a common
procedure is to do a Nudged Elastic Band (NEB) calculation. Where all the
intermediate replicas are created and total energy calculations are done on
each replica. This calculation will help us to know the entire trajectory of
the vacancies. But this is also to be kept in mind that these calculations are
very expensive in time.
133
xyz
vacancy
xyz
movementof the vacancy
Figure 5.26: The real picture of the movement of the vacancy
5.10.3 Homothetic clusters
There is a possibility to account for the surface effect in a more analytical
and quantitative manner, if we can create the clusters of different sizes with
the same family of surfaces. Clusters of different sizes with same surfaces
are called homothetic clusters. If we can put one cluster inside the other one
they will make something like concentric clusters. We have already succeeded
in making homothetic clusters of three different sizes but calculations with
them took us out of the time limit of the present thesis. We believe similar
calculations as done within the framework of this thesis will yield interesting
quantitative facts about th surface effect in the nanoclusters.
134
xy
z
xy
z
xy
z
xy
z
Figure 5.27: Homothetic clusters made from boxes of (a)64 (b)216 (c)512(d)1000 silicon atoms.
135
Chapter 6
Final words and Perspectives
In the quest to study the charged defects of silicon we have managed to
touch upon and dig into many different related topics in this branch of con-
densed matter physics. In terms of field study, we have gathered knowledge
about almost all the techniques in the DFT framework, that scientists use to
calculate various physical quantities regarding the charged defects. In that
kind of a study we have summarized the advantages and disadvantages of
each method. Indeed PBC methods are good to simulate a bulk, but the
problem of image defect interactions is something which we cannot get rid of
anyhow, barring the use of some correction factors. In the main part of this
research we have tried to find an approach which would be different from the
PBC formalisms. We have felt the need of not using the background com-
pensating charge in the PBC formalism for the calculation of the charged
defects. The backbone of this research is the DFT method as all the total
energy calculations are done with the wavelet-based DFT code - BigDFT.
Hence a brief yet detailed description of the DFT framework is included in
the thesis. The important concepts about the basis sets, pseuopotentials and
the exchange correlation functionals are also discussed in here. In the study
with the HGH pseudopotentials, we have tried to show the reliability of those
pseudos with the inclusion of the NLCC. We have shown of course that with
the NLCC implemented within the HGH pseudopotentials, we can use them
for very accurate calculations, as comparable to the all electron calculations.
136
This is a major step towards the pseudopotential calculations as it takes us
towards the ultimate aim, which is to use a complexity-less scheme for accu-
rate research. The main part of this thesis as I have already mentioned was
to find a novel way to calculate the formation energy of the charged defects.
We have used nanoclusters to simulate the defects and have found that the
formation energy thus found are in accordance with whatever experimental
values we have. We have put forward the phenomenon of the trapped charges
near the vacancy in the nanoclusters which we believe affects the geometry
and the energy states of the defect. Keeping that in mind we have given here
an analytical electrostatic model to simulate the charge distribution in the
nanoclusters. We have also done various numerical analyses to provide ample
proof for the electrostatic model of the system. Although an admitted fact is
that the PBC method employs less atoms than our FBC method but in our
method we have dealt away with the correction factors needed otherwise for
the PBC method. In future these analyses with the correct electrostatics can
be used for other systems and of course for various other studies like the mi-
gration energy and diffusion of defects. We can study the effect of charge in
the substitutional and interstitial defects in silicon with this cluster method.
In principle the success and validity of this method would ensure further
application of this to study all other point defects in this new method. It
would be very useful to know about the energetics and the geometry of such
defects in nanoclusters. There is of course a lot of room to study magnetic
defects as well. Nanoclusters which are magnetic can also be studies within
similar formalism. The effect of the trapped charges and the surfaces in the
magnetic systems can be extremely interesting to look for. We can look for
the various effects of the dilute doping defects and the resultant magnetism in
the metal oxide systems with this method. This method will provide us with
an alternative way to the model more frequently used hamiltonian methods.
In a nutshell we can say a new method to study the energetics of the defect
will open up channels of research with defects in all possible environment.
With the recent development to include the applicability of the Wannier
functions in BigDFT, we can be hopeful of using this cluster approach to
137
carry out somewhat an order-N method, in terms of localization regions,
to study some embedded systems. As for example we can talk about the
molecules on surfaces which are semi-infinite systems, or even the charged
defects in bulk(infinite systems). The correct electrostatics of the cluster ap-
proach makes it transferable to the embedded system calculations. Since the
wavelets are localized, for each orbital a region of localization can be defined
. All the operations can be carried out as applications of the Hamiltonian
in the given orbital. Hence methods and theoretical tools can be developed
based on order N method to perform calculation of an atomic system embed-
ded by a tight-binding model based on Wannier functions. If the Wannier
orbitals of Si bulk can be calculated beforehand, then, using tight binding
methods on Wannier, one can define a hamiltonian and a self-consistent field
to calculate an infinite system of Si bulk. Following steps would include the
processes to deal with the defect in the embedded environment along with
the relaxation of the Wannier orbitals in a cluster centered around a vacancy.
It will be feasible with this mixed scheme (wavelet + tight-bindings) to de-
scribe properly charged trapped states even if dimension of the system exceed
many nanometers. It is worth mentioning here that wavelets being localized
themselves make it possible to do locally all operations as applications of the
Hamiltonian. It terms of technical complexities wavelets are advantageous
than plane waves in this particular aspect. Hence we can consider the clus-
ter approach, as we have described in the larger part of this thesis, as the
first step in the field of an order-N method to study the embedded systems.
With the electrostatics and the energetics known for the cluster method the
following research procedures will be immensely facilitated .
138
Bibliography
[1] D. A. Faux, J. R. Downes, E. P. O’Reilly, J. Appl. Phys., 82, 3754–3762
(1997).
[2] G. A. Baraff, E. O. Kane, M. Schluter, Phys. Rev. Lett., 43, 956–959
(1979).
[3] G. D. Watkins, J. W. Corbett, Phys. Rev., 138, A543–A555 (1965).
[4] G. D. Watkins, J. R. Troxell, Phys. Rev. Lett., 44, 593–596 (1980).
[5] J. C. Bourgoin, J. W. Corbett, Rad. Effects, 36, 157–188 (1978).
[6] R. J. Needs, J. Phys. Condens. Matter 11, 10437–10450 (1999).
[7] W.-K. Leung, R. J. Needs, G. Rajagopal, S. Itoh, S. Ihara, Phys. Rev.
Lett., 83, 2351–2354 (1999).
[8] S. J. Clark, G. J. Ackland, Phys. Rev. B, 56, 47–50 (1997).
[9] V. V. Lukianitsa, Semiconductors, 37, 404–413 (2003).
[10] G. M. Lopez, V. Fiorentini, Phys. Rev. B, 69, 155206–155213 (2004).
[11] A. H. Kalma, J. C. Corelli, Phys. Rev., 173, 734–745 (1968).
[12] L. J. Cheng, J. C. Corelli, J. W. Corbett, G. D. Watkins, Phys. Rev.,
152, 761–774 (1966).
[13] S. Goedecker, T. Deutsch, L. Billard, Phys. Rev. Lett., 88,
235501–235504 (2002).
139
[14] I. Lazanu, S. Lazanu, Phys. Scr., 74, 201–207 (2006).
[15] D.Caliste and P.Pochet, Phys. Rev. Lett. 97, 135901 (2006).
[16] D.Caliste, K.Z.Rushchanskii, and P.Pochet , Appl. Phys. Lett. 98,
031908 (2011).
[17] P. Hohenberg and W. Kohn, Phys. Rev. 136. B864 (1964).
[18] W. Kohn, Reviews of Modern Physics 71 nb. 5 (1999).
[19] W. Kohn and L. J. Sham, Phys. Rev 140, A1133 (1965).
[20] C.-O. Almbladh, U. von Barth Phys., Rev. B 31, nb. 6 (1984).
[21] J. Chen, J. B. Krieger, Y. Li, J.G. Iafrate, Phys. Rev. A 54 nb. 5 (1996).
[22] L Fan, T.J. Ziegler J. Chem. Phys. 95, 7401 (1991).
[23] J. P. Perdew, A. Zunger, Phys. Rev. B 23, nb. 10 (1981).
[24] S. Kurth, J. P. Perdew, Phys. Rev. B 59 nb. 16 (1999)
[25] J. P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. 77, 3865
(1996).
[26] M. Ernzerhof and G. E. Scuseria, J. Chem. Phys. 110, 5029 (1999).
[27] J. P. Perdew and Y. Wang, Phys. Rev. B 45, 13244 (1992).
[28] J. P. Perdew, S. Kurth, A. Zupan, and P. Blaha, Phys. Rev. Lett. 82,
2544 (1999).
[29] von Barth, U., Physica Scripta. Vol. T109, 9–39, (2004).
[30] Muller, J. E., Jones, R. O. and Harris, J., J. Chem. Phys. 79,
1874(1983).
[31] Terakura, K., Oguchi, T., Williams, A. R. and Kubler, J., Phys. Rev. B
30, 4734 (1984).
140
[32] Wang, C. S., Klein, B. M. and Krakauer, H., Phys. Rev. Lett. 54, 1852
(1985).
[33] Godby, R. W. and Needs, R. J., Phys. Rev. Lett. 62, 1169 (1989).
[34] Parr, R. G. and Yang, W., Density-Functional Theory of Atoms and
Molecules,(Oxford University Press, New York 1989)
[35] von Barth, U., Chem. Scripta 26, 449 (1986).
[36] J. Tao, J. P. Perdew, V. N. Staroverov, and G. E. Scuseria, Phys. Rev.
Lett. 91, 146401 (2003).
[37] J. P. Perdew, J. Tao, V. N. Staroverov, and G. E. Scuseria, J. Chem.
Phys. 120, 6898 (2004).
[38] J. Harris and R. O. Jones, J. Phys. F 4, 1170 (1974).
[39] A. Gorling and M. Levy, J. Chem. Phys. 106, 2675 (1997).
[40] A. D. Becke, J. Chem. Phys. 98, 5648 (1998).
[41] E.R.Batista, J.Heyd, R.G.Hennig, B.P.Uberuaga, R.L.Martin,
G.E.Scuseria, C.J.Umrigar and J.W.Wilkins Phys. Rev.B 74, 121102
R(2006).
[42] J. Heyd, G. E. Scuseria, and M. Ernzerhof, J. Chem. Phys. 118, 8207
(2003); J. Heyd and G. E. Scuseria, ibid. 120, 7274[U+0351](2004) ; J.
Heyd and G. E. Scuseria, ibid. 121, 1187 (2004).
[43] L. Genovese, A. Neelov, S. Goedecker, T. Deutsch, S. Ghasemi, A. Wil-
land, D. Caliste, O. Zilberberg, M. Rayson, A. Bergman,and R. Schneider
, J. Chem. Phys. 129, 014109 2008
[44] D.R. Hamann, M. Schluter, and C. Chiang, Phys. Rev. Lett. 43, 1494
(1979).
[45] G.B. Bachelet, D.R. Hamann, and M. Schluter, Phys. Rev. B 26, 4199
(1982).
141
[46] G. Kerker, J. Phys. C 13, L189 (1980).
[47] D. Vanderbilt, Phys. Rev. B 32, 8412 (1985).
[48] N. Troullier and J.L. Martins, Phys. Rev. B 43, 1993 (1991).
[49] S. Goedecker, M. Teter, and J. Hutter, Phys. Rev. B 54, 1703 (1996).
[50] S. Goedecker , K. Maschke Phys. Rev. A 45, 88–93 (1992) .
[51] C. Hartwigsen, S. Goedecker, and J. Hutter, Phys. Rev. B 58, 3641
(1998).
[52] A.M. Rappe, K.M. Rabe, E. Kaxiras, and J.D. Joannopoulos, Phys.
Rev. B 41, 1227(1990).
[53] L. Kleinman and D.M. Bylander, Phys. Rev. Lett. 48, 1425 (1982).
[54] D. Vanderbilt, Phys. Rev. B 41, 7892 (1990).
[55] K. Laasonen, A. Pasquarello, R. Car, C. Lee, and D. Vanderbilt, Phys.
Rev. B 47,10142 (1993).
[56] G.P. Francis, and M. C. Payne, J. Phys. Condens. Matter 17, 1643,
(1990).
[57] T. L. Beck, Rev. Mod. Phys. 72, 1041 2000.
[58] J. E. Pask, B. M. Klein, C. Y. Fong, and P. A. Sterne, Phys. Rev. B
59,12352 1999.
[59] J. J. Mortensen, L. B. Hansen, and K. W. Jacobsen, Phys. Rev. B
71,035109 2005.
[60] M.C.Payne, M.P.Teter,D.C.Allan,T.A.Arias,and J.D.Joannopoulos,
Rev. Mod. Phys. 64, 1045(1992).
[61] A.M.Rappe,J.D.Joannopoulos, and P.A.Bash, J. Am. Chem. Soc. 114,
6466(1992).
142
[62] M.Leslie and M.J.Gillian, J. Phys. C 18, 973 (1985).
[63] G. D. Watkins Material science in semiconductor processing 3 227-235
(2000)
[64] G. Feher Phys. Rev. 103 (3): 834–835, (1956)
[65] R. O. Simmons and R. W. Balluffi Phys. Rev. 117, 52–61 (1960).
[66] G.Makov and M.C.Payne, Phys. Rev. B. 51, 7(1995).
[67] S.W.de Leeuw,J.W¿Perram, and E.R.Smith, Proc. R. Soc. London Ser.
A373, 27 (1980).
[68] R.M.Martin, Phys. Rev. B 9,1998 (1974).
[69] Peter A.Schultz, Phys. rev. Let. 84, 9 (2000).
[70] Peter A.Schultz, Phys. rev. B 60, 1551 (1999).
[71] M.Leslie and M.J.Gillian, J. Phys. C 18, 973 (1985).
[72] C.Freysoldt, J.Neugebauer,C.G.Van de Walle, Phys. rev. Let. 102,
016402 (2009).
[73] C. W. M. Castleton, A. Hoglund, and S. Mirbt, Phys. Rev.B 73, 035215
(2006).
[74] J. Shim, E.K. Lee, Y. J. Lee, and R. M. Nieminen, Phys. Rev. B 71,
035206 (2005).
[75] A. F. Wright and N. A. Modine, Phys. Rev. B 74, 235209 (2006).
[76] I. Dabo, B.Kozinsky, N.E.Singh-Miller, and N.Marzari, Phys. Rev. B
77, 115139 (2008)
[77] L. A. Curtiss, K. Raghavachari, P. C. Redfern, and J. A. Pople, J. Chem.
Phys. 106, 1063 1997.
[78] G. Sun, J. Kurti, P. Rajczy, M. Kertesz, J. Hafner, and G. Kresse, J.
Mol. Struct.: THEOCHEM 624, 37 2003.
143
[79] S. B. Zhang, S. H. Wei, and Z. Zunger, Phys. Rev. B 63, 075205 2001.
[80] P. E. Blochl, Phys. Rev. B 50, 17953 1994.
[81] G. Kresse and J. Hafner, Phys. Rev. B 47, 558 1993.
[82] B. Hammer, L. B. Hansen, and J. K. Nørskov, Phys. Rev. B 59, 7413
1999.
[83] E. L. Briggs, D. J. Sullivan, and J. Bernholc, Phys. Rev. B 54, 14362
1996.
[84] L. A. Curtiss, K. Raghavachari, G. W. Trucks, and J. A. Pople, J. Chem.
Phys. 94, 7221 1991.
[85] J. Paier, R. Hirschl, M. Marsman, and G. Kresse, J. Chem. Phys. 122,
234102 2005.
[86] S. Goedecker, M. Teter and J. Hutter, Phys. Rev. B 54, 3 1996.
[87] S.G.Louie, S.Froyen and M.L.Cohen, Phys. Rev.B 26, 4 1982
[88] A. Zunger, Phys. Rev. B 22, 649 1980
[89] D. R. Hamann, M. Schluter, and C. Chiang, Phys.Rev. Lett. 43, 1494
1979
[90] M. T. Yin and M. L. Cohen, Phys. Rev. Lett. 45, 1004 1980
[91] R. Fournier, L. Boroukhovskaia, Theor. Chem. Acc. 112: 1–6 2004
[92] P. Rinke, C. G. V. de Walle, and M. Scheffler, Phys. Rev. Lett. 102,
026402 (2009).
[93] R. Nieminen, Modell. Simul. Mater. Sci. Eng. 17, 084001 (2009).
[94] A. F. Wright and N. A. Modine, Phys. Rev. B 74, 235209 (2006).
[95] J. R. Chelikowsky, N. Troullier, and Y. Saad, Phys. Rev. Lett. 72, 1240
1994[U+0352].
144
[96] E. L. Briggs, D. J. Sullivan, and J. Bernholc, Phys. Rev. B 52, R5471
1995.
[97] S. Ogut, H. Kim, J. R. Chelikowsky Phys. rev. B. (Rapid Comm.) 56,
18 1997.
[98] C.G. Van de Walle, J. Neugebauer, J. Appl. Phys. 95, 8 2004.
[99] R. S. Mulliken, J. Chem. Phys. 23, 1833 (1955).
[100] D.V. Makhov and L.J. Lewis, Phys. Rev. B 72, 073306 2005.
[101] S. Ogut and J. R. Chelikowsky Phys. Rev. Lett. 83, 19 (1999)
[102] M. Saito and A. Oshiyama, Phys. Rev. Lett. 73, 866 1994.
[103] P.Pulay, Chem. Phys. Lett. 73, 2, 393-398 (1990)
145
Acknowledgments
I am extremely grateful to Nanoscience Foundation who funded this project
for the first three years. All the staff members of the Foundation were very
helpful with the administrative proceedings like the VISA for me and my
wife, the reimbursements of the transport expenses, paper work for the tax
payment etc. I would like to thank them for such a support and wish them
all the best to carry on with the wonderful work of helping researchers from
all around the world.
The last four months of my PhD thesis was funded by CEA and I have
no words to thank for their kind consideration. I also thank all the adminis-
trative staff members to help me with all the necessary paper works needed
for such an extension.
To start with I would like to thank Dr. Damien Caliste, my PhD super-
visor, to have been so helpful. Through all the small and silly mistakes to
the biggest blunders he kept his support and helped me in all possible ways
to finish my work. There are so many important things that I learnt from
him. From computer techniques to concepts of physics discussions with him
always ended in my gaining some new knowledge. It was nice to know with
him the different aspects of research. I can’t thank him enough for his help
in difficult times of the research work or even the administrative hustles. I
thank him for his patience to deal with my innumerable mistakes in the re-
search work. It was very nice to know him and his family and I would be
ever thankful for his all round support.
I would take this opportunity to thank Dr. Thierry Deutsch, my director
to thesis, for letting me have the opportunity to work on this interesting
topic. I got to learn a lot about various aspects of physics and computer
science because of many important discussions with him. I am also thankful
for his kind consideration in letting me have an extension of my PhD con-
tract, which otherwise is very difficult to have as I gather. There had been
146
strained patches in between but I would take the positive out of it to thank
you for letting me know the hardships of research and I guess I have learnt
many important things to do a good research. Thanks a lot for being there.
I would like to thank Mr. Alex Willand and Dr. Stefan Goedecker for
collaborating on a project with me. Discussion through email with them had
been of great help in the research.
In my laboratory there are so many people to whom I was thankful for
various reasons. A special thanks to Dr. Luigi Genovese for some delightful
discussions regarding the collaborative work we did. It was great to learn so
many different aspects of physics and computer science from you. I would
like to thank Dr. Pascal Pochet, Dr. Yann-Michel Niquet, Dr. Frederic
Lancon and Dr. Brice Videau, discussions with whom helped me to learn a
lot about different things in physics and computer science. They had been
very understanding and patient to all my questions irrespective of their being
related to science or something else.
I had wonderful friends in the lab in the past three and half years. Each
of them helped me in all possible ways and words fail me when I wish to
thank them for being there with me. I would like to thank Thomas Jour-
dan, at whose party for the first time I got to know my friends very well. I
am thankful to Emanuel Arras, who helped me with my initial proceedings
in research with all sorts of things in physics and computers. I am equally
thankful to Martin Presson and Margaret Gabriel Presson for their wonderful
companionship. A very special thanks to Ivetta Slipukhina for being there as
a constant support encouraging to be positive in times when I was down. I
would like to thank Dulce Camacho for being such a wonderful friend, sharing
innumerable coffees and chats of all kinds. She has also been a great source of
inspiration. I would also thank Hector Mera for his tremendous support for
me and my work and for the wonderful times that we got to spend together.
His insight to science and inspiration to think differently have helped me a lot
in my journey. To my friend Cornelia Metzig I would be extremely grateful
147
as well for she helped me in many ways. Discussion with her about physics,
science in general, society, philosophy and many other things made my stay
in the lab very enjoyable. I would also thank Eduardo Machado Charry for
helping me with many small but important things regarding my thesis. His
presence always lights up the place. I would also like to thank Irinia Georgina
Groza for being a lovely inspiring and encouraging friend. My gratitude goes
for Bhaaraathi Natarajan, discussions with whom is something which I had
always enjoyed.
During my stay in France, there is a family to whom I will never be able
to say how grateful I am to have known them - Genvieve Stella Moriceau
and Hubert Moriceau. They have love us like their own and have helped us
in every possible way to make our lives safe, healthy and enjoyable. They
had been the greatest strength and I would cherish every moment of those
I spent with them. With them I got to know so much about the culture in
France and rest of Europe. They had been the haven where I would run every
time I am in some trouble and of course they would help me with whatever
they could. It’s been an honor to have known them so closely.
In these three and half years I became friends with so many people. Of
them all the one who became almost a part of me was Akash Chakraborty.
Now, to look back in the times that we spent together, I feel it was a part of
myself that was moving around with me. We drank and ate all that we could,
went to places together, watched and discussed about innumerable movies,
talked about all that we could find under the sun and most of all had the
best possible time between us. In days of hard work and distress his support
was like the protective imperceptible ether. Can I thank him at all for being
a part of my senses?
Friends like Soumen Mondal, Anita Sarkar, Dibyendu Hazra, Subhadeep
Dutta, Kalpana Mondal became more than family. Innumerable discussions
regarding various subjects that spanned our work, culture, personal life etc.
made our days in Grenoble so beautiful. A very special thanks to Soumen
148
Mondal for helping in with his ambidextrous power of mending household
things, giving opinions in research and entrepreneurship way of life. I would
also like to extend my gratitude to friends like Ayan Kumar and Kuheli Ban-
dopadhyay, Chandan Bera, Sutirtha and Dipanwita Mukherjee, Arijit and
Satarupa Roy, Abhijit Ghosh, Anupam Kundu, Sreeram Venkatpathy, Shilpi
Sreeram and Pragyan, Roopak Sinha, Arpita Sinha, Dhrovi and Diu for be-
ing there and making the days in Grenoble happy and enjoyable for us. I
would like to thank Biplab Mondal for lending me his camera for most of my
stay here in Grenoble. It has been really like a family I wish it stays like that
forever. Ananda Sankar Basu and Priyadarshini Chatterjee deserve a special
bunch of thanks for being almost always the driving force of this extended
family of ours. It’s been a privilege to know you all.
This thesis is probably an important milestone in this journey and I guess
it had started under the guidance of the teachers who I think I can never
properly thank. Dr. Ananda Dasgupta is one of those teachers to whom I
feel I could go at any point of time to ask any question regarding physics. It
is actually a great feeling of confidence to have when day in day out some-
one is struggling to reach an intellectual height. I would never be able to
thank him enough. Prof. Albert Gomes had been an infinite source of in-
spiration with his intense dedication for students and science. I would like
to express my gratitude for his being there with me all those times when I
needed inspiration to remain dedicated, focused and honest. I would like to
thank Dr. Dipankar Sengupta, Dr. Indranath Chaudhury, Dr. Tapati Dutta,
Dr. Shibaji Banerjee and Dr. Suman Bandopadhyay who had taught me so
many things in science and have always inspired. I am extremely thankful
to Prof. D.G. Kanhere who taught me so many different important things
in physics and computer science. This PhD wouldn’t have been possible if
it wasn’t him who inspired me to come here and work. He has been a great
source of inspiration to all the students as he has been for me. I would also
thank Dr. Anjali Kshirsagar, with whom I managed to learn various aspect
of computational physics and physics in general. I would also like to thank
Mr. Pinaki Mitra and Mr. Nimai Bhattacharya who always have preached
149
to me the reason to use life for the quest of higher knowledge of all kinds.
I am grateful to Mrs. Jayati Acharya whose good wishes and relentless en-
couragement along with her sublime music lessons helped me a lot to grow
as a human being and sustain under tremendous pressure.
Honestly it was my mother Mrs. Chhabi Deb who dreamt her son becoming a
scientist one day. The constant motivating factor above all was my mother’s
dream about this day. I cannot thank her enough for having had that dream
and pushing me hard enough so that I could reach where I am now. My
father Mr. Amit Krishna Deb, with his utmost honesty and sincerity to life
has forced me to follow the path of honesty and integrity even when I could
have drifted. It’s their sincere love and care for me which made me worthy
of whatever I have achieved till date. I am glad that I have a brother like
Mr. Ayan Krishna Deb, who loved me like his own self and even in the most
difficult of our days his love was as true as the sun. I would like to thank
Mr. Asesh Krishna Deb, my eldest cousin whose constant support always
provided me with tremendous confidence. I would also like to thank my all
other family members and all other friends whose love and affection had been
a constant source of inspiration.
I would like to take this opportunity to express my gratitude for the late
Mr. Samir Mukherjee who taught me through the first days of science and
mathematics. I would like to thank my oldest school friends Joy Banerjee
and Angshuman Samadder who had been with me though thick and thin and
who have inspired me through examples all these days.
Among the friends whose constant help saw me through the initial days
of the journey with physics, I would like to thank Arya Dhar, Amit Kumar
Pal, Sayan Chakraborty, Beas Roy, and Sumitra Raymoulik. They taught
me all that was in syllabus, made notes for me to pass, even introduced me
to various new subjects. It is great to know them so closely. I would like
to thank my friends Aditya Rotti, Jaydeep Belapure, Sumati Patil, Deep-
ashri Saraf, Narendranath Patra, Sayan Mondal, Anirban Pal and Saurav
150
Roy because they always believed I could reach this far. I would take this
opportunity to thank our Ajji Mrs. Subha Joshi who have always loved me
so much and have inspired me immensely to thrive harder for my goal in life.
I would also thank my friend Maitra, whose constant adulation and indul-
gence made me believe that I can be a creative individual as well. Her
support, encouragement and friendship is something which one can always
take pride in. Her honesty and integrity to life itself have been a great source
of inspiration to me and my family.
In the end, I would like to thank the person, thanking whom seems ab-
solutely insignificant, for it is Mrs. Maitreyee Mookherjee, who essentially
lived a life completely for my cause and sake. That she gave up her own
career to be with me through the times I needed her more than ever, that
she had faith in me when even I had lost faith in life, that she loved me even
when I failed her, are reasons enough for not thanking her. One cannot thank
someone who has done so much. Her dedication and love for me has been
my energy so far. I couldn’t have made it anywhere without her constant
support, encouragement and love. I could have written pages for her but it
would mean the same - that I wish she stays there with me forever.
151